[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;
	bool overlap = false;

	do {
		/*
		 * 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
		 */
		pthread_mutex_lock(&overlap_check);
		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;

			overlap = in_flight_overlap(&td->io_u_all, io_u);
			if (overlap) {
				pthread_mutex_unlock(&overlap_check);
				break;
			}
		}
	} while (overlap);
}

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;

	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;
		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;
	}

	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_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;
	16	bool overlap = false;
	17
	18	do {
	19	/*
	20	* Allow only one thread to check for overlap at a
	21	* time to prevent two threads from thinking the coast
	22	* is clear and then submitting IOs that overlap with
	23	* each other
	24	*/
	25	pthread_mutex_lock(&overlap_check);
	26	for_each_td(td, i) {
	27	if (td->runstate <= TD_SETTING_UP \|\|
	28	td->runstate >= TD_FINISHING \|\|
	29	!td->o.serialize_overlap \|\|
	30	td->o.io_submit_mode != IO_MODE_OFFLOAD)
	31	continue;
	32
	33	overlap = in_flight_overlap(&td->io_u_all, io_u);
	34	if (overlap) {
	35	pthread_mutex_unlock(&overlap_check);
	36	break;
	37	}
	38	}
	39	} while (overlap);
	40	}
	41
155f2f02 JA	42	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	43	struct workqueue_work *work)
40511301 JA	44	{
	45	struct io_u *io_u = container_of(work, struct io_u, work);
	46	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	47	struct thread_data *td = sw->priv;
40511301 JA	48	int ret;
40511301 JA	49
c06379a6 VF	50	if (td->o.serialize_overlap)
	51	check_overlap(io_u);
	52
40511301 JA	53	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
40511301 JA	54
1651e431	55	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	56
	57	td->cur_depth++;
	58
	59	do {
	60	ret = td_io_queue(td, io_u);
	61	if (ret != FIO_Q_BUSY)
	62	break;
	63	ret = io_u_queued_complete(td, 1);
	64	if (ret > 0)
	65	td->cur_depth -= ret;
1651e431	66	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	67	} while (1);
	68
	69	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	70
	71	io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
	72
	73	if (ret == FIO_Q_COMPLETED)
	74	td->cur_depth--;
	75	else if (ret == FIO_Q_QUEUED) {
	76	unsigned int min_evts;
	77
	78	if (td->o.iodepth == 1)
	79	min_evts = 1;
	80	else
	81	min_evts = 0;
	82
	83	ret = io_u_queued_complete(td, min_evts);
	84	if (ret > 0)
	85	td->cur_depth -= ret;
40511301	86	}
155f2f02 JA	87
155f2f02 JA	88	return 0;
40511301 JA	89	}
	90
	91	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	92	{
b86ad8f1	93	struct thread_data *td = sw->priv;
40511301 JA	94
	95	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	96	return true;
	97
	98	return false;
	99	}
	100
	101	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	102	{
b86ad8f1	103	struct thread_data *td = sw->priv;
40511301 JA	104	int ret;
	105
	106	ret = io_u_quiesce(td);
	107	if (ret > 0)
	108	td->cur_depth -= ret;
	109	}
	110
	111	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	112	{
	113	struct thread_data *td;
	114
	115	td = calloc(1, sizeof(*td));
b86ad8f1	116	sw->priv = td;
40511301 JA	117	return 0;
	118	}
	119
	120	static void io_workqueue_free_fn(struct submit_worker *sw)
	121	{
b86ad8f1 CB	122	free(sw->priv);
b86ad8f1 CB	123	sw->priv = NULL;
40511301 JA	124	}
	125
	126	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	127	{
	128	struct thread_data *parent = sw->wq->td;
b86ad8f1	129	struct thread_data *td = sw->priv;
40511301 JA	130
	131	memcpy(&td->o, &parent->o, sizeof(td->o));
	132	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	133	td->o.uid = td->o.gid = -1U;
	134	dup_files(td, parent);
	135	td->eo = parent->eo;
	136	fio_options_mem_dupe(td);
	137
	138	if (ioengine_load(td))
	139	goto err;
	140
40511301 JA	141	td->pid = gettid();
	142
	143	INIT_FLIST_HEAD(&td->io_log_list);
	144	INIT_FLIST_HEAD(&td->io_hist_list);
	145	INIT_FLIST_HEAD(&td->verify_list);
	146	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	147	td->io_hist_tree = RB_ROOT;
	148
	149	td->o.iodepth = 1;
	150	if (td_io_init(td))
	151	goto err_io_init;
	152
3aea75b1	153	set_epoch_time(td, td->o.log_unix_epoch);
40511301 JA	154	fio_getrusage(&td->ru_start);
	155	clear_io_state(td, 1);
	156
	157	td_set_runstate(td, TD_RUNNING);
b7aae4ba	158	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	159	td->parent = parent;
	160	return 0;
	161
	162	err_io_init:
	163	close_ioengine(td);
	164	err:
	165	return 1;
	166
	167	}
	168
	169	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	170	unsigned int *sum_cnt)
	171	{
b86ad8f1	172	struct thread_data *td = sw->priv;
40511301 JA	173
	174	(*sum_cnt)++;
	175	sum_thread_stats(&sw->wq->td->ts, &td->ts, *sum_cnt == 1);
	176
	177	fio_options_free(td);
	178	close_and_free_files(td);
	179	if (td->io_ops)
	180	close_ioengine(td);
	181	td_set_runstate(td, TD_EXITED);
	182	}
	183
	184	#ifdef CONFIG_SFAA
	185	static void sum_val(uint64_t dst, uint64_t src)
	186	{
	187	if (*src) {
	188	__sync_fetch_and_add(dst, *src);
	189	*src = 0;
	190	}
	191	}
	192	#else
	193	static void sum_val(uint64_t dst, uint64_t src)
	194	{
	195	if (*src) {
	196	dst += src;
	197	*src = 0;
	198	}
	199	}
	200	#endif
	201
	202	static void pthread_double_unlock(pthread_mutex_t *lock1,
	203	pthread_mutex_t *lock2)
	204	{
	205	#ifndef CONFIG_SFAA
	206	pthread_mutex_unlock(lock1);
	207	pthread_mutex_unlock(lock2);
	208	#endif
	209	}
	210
	211	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	212	{
	213	#ifndef CONFIG_SFAA
	214	if (lock1 < lock2) {
	215	pthread_mutex_lock(lock1);
	216	pthread_mutex_lock(lock2);
	217	} else {
	218	pthread_mutex_lock(lock2);
	219	pthread_mutex_lock(lock1);
	220	}
	221	#endif
	222	}
	223
	224	static void sum_ddir(struct thread_data dst, struct thread_data src,
	225	enum fio_ddir ddir)
	226	{
	227	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	228
	229	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	230	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	231	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	232	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	233	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
	234
	235	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	236	}
237
238	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
239	{
b86ad8f1	240	struct thread_data *src = sw->priv;
40511301 JA	241	struct thread_data *dst = sw->wq->td;
	242
	243	if (td_read(src))
	244	sum_ddir(dst, src, DDIR_READ);
	245	if (td_write(src))
	246	sum_ddir(dst, src, DDIR_WRITE);
	247	if (td_trim(src))
	248	sum_ddir(dst, src, DDIR_TRIM);
	249
	250	}
	251
103b174e	252	static struct workqueue_ops rated_wq_ops = {
40511301 JA	253	.fn = io_workqueue_fn,
	254	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	255	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	256	.update_acct_fn = io_workqueue_update_acct_fn,
	257	.alloc_worker_fn = io_workqueue_alloc_fn,
	258	.free_worker_fn = io_workqueue_free_fn,
	259	.init_worker_fn = io_workqueue_init_worker_fn,
	260	.exit_worker_fn = io_workqueue_exit_worker_fn,
	261	};
103b174e	262
24660963	263	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	264	{
	265	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	266	return 0;
	267
24660963	268	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	269	}
	270
	271	void rate_submit_exit(struct thread_data *td)
	272	{
	273	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	274	return;
	275
	276	workqueue_exit(&td->io_wq);
	277	}