[fio.git] / rate-submit.c

/*
 * Rated submission helpers
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include <assert.h>
#include <errno.h>
#include <pthread.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);
	if (fio_unlikely(res != 0)) {
		log_err("failed to lock overlap check mutex, err: %i:%s", errno, strerror(errno));
		abort();
	}

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);
		if (fio_unlikely(res != 0)) {
			log_err("failed to unlock overlap check mutex, err: %i:%s", errno, strerror(errno));
			abort();
		}
		res = pthread_mutex_lock(&overlap_check);
		if (fio_unlikely(res != 0)) {
			log_err("failed to lock overlap check mutex, err: %i:%s", errno, strerror(errno));
			abort();
		}
		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);

	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]);

	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>
83276370 DP	8	#include <errno.h>
	9	#include <pthread.h>
	10
40511301	11	#include "fio.h"
618ee94c	12	#include "ioengines.h"
40511301	13	#include "lib/getrusage.h"
24660963	14	#include "rate-submit.h"
40511301	15
c06379a6 VF	16	static void check_overlap(struct io_u *io_u)
c06379a6 VF	17	{
a38c70a8	18	int i, res;
c06379a6	19	struct thread_data *td;
c06379a6	20
c76b661c BVA	21	/*
	22	* Allow only one thread to check for overlap at a time to prevent two
	23	* threads from thinking the coast is clear and then submitting IOs
	24	* that overlap with each other.
	25	*
	26	* If an overlap is found, release the lock and re-acquire it before
	27	* checking again to give other threads a chance to make progress.
	28	*
	29	* If no overlap is found, release the lock when the io_u's
	30	* IO_U_F_FLIGHT flag is set so that this io_u can be checked by other
	31	* threads as they assess overlap.
	32	*/
a38c70a8	33	res = pthread_mutex_lock(&overlap_check);
83276370 DP	34	if (fio_unlikely(res != 0)) {
	35	log_err("failed to lock overlap check mutex, err: %i:%s", errno, strerror(errno));
	36	abort();
	37	}
c76b661c BVA	38
	39	retry:
	40	for_each_td(td, i) {
	41	if (td->runstate <= TD_SETTING_UP \|\|
	42	td->runstate >= TD_FINISHING \|\|
	43	!td->o.serialize_overlap \|\|
	44	td->o.io_submit_mode != IO_MODE_OFFLOAD)
	45	continue;
c06379a6	46
c76b661c BVA	47	if (!in_flight_overlap(&td->io_u_all, io_u))
	48	continue;
	49
a38c70a8	50	res = pthread_mutex_unlock(&overlap_check);
83276370 DP	51	if (fio_unlikely(res != 0)) {
	52	log_err("failed to unlock overlap check mutex, err: %i:%s", errno, strerror(errno));
	53	abort();
	54	}
a38c70a8	55	res = pthread_mutex_lock(&overlap_check);
83276370 DP	56	if (fio_unlikely(res != 0)) {
	57	log_err("failed to lock overlap check mutex, err: %i:%s", errno, strerror(errno));
	58	abort();
	59	}
c76b661c BVA	60	goto retry;
c76b661c BVA	61	}
c06379a6 VF	62	}
c06379a6 VF	63
155f2f02 JA	64	static int io_workqueue_fn(struct submit_worker *sw,
155f2f02 JA	65	struct workqueue_work *work)
40511301 JA	66	{
	67	struct io_u *io_u = container_of(work, struct io_u, work);
	68	const enum fio_ddir ddir = io_u->ddir;
b86ad8f1	69	struct thread_data *td = sw->priv;
d28174f0	70	int ret, error;
40511301	71
c06379a6 VF	72	if (td->o.serialize_overlap)
	73	check_overlap(io_u);
	74
40511301 JA	75	dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
40511301 JA	76
1651e431	77	io_u_set(td, io_u, IO_U_F_NO_FILE_PUT);
40511301 JA	78
	79	td->cur_depth++;
	80
	81	do {
	82	ret = td_io_queue(td, io_u);
	83	if (ret != FIO_Q_BUSY)
	84	break;
	85	ret = io_u_queued_complete(td, 1);
	86	if (ret > 0)
	87	td->cur_depth -= ret;
d28174f0 JA	88	else if (ret < 0)
d28174f0 JA	89	break;
1651e431	90	io_u_clear(td, io_u, IO_U_F_FLIGHT);
40511301 JA	91	} while (1);
	92
	93	dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
	94
d28174f0	95	error = io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
40511301 JA	96
	97	if (ret == FIO_Q_COMPLETED)
	98	td->cur_depth--;
	99	else if (ret == FIO_Q_QUEUED) {
	100	unsigned int min_evts;
	101
	102	if (td->o.iodepth == 1)
	103	min_evts = 1;
	104	else
	105	min_evts = 0;
	106
	107	ret = io_u_queued_complete(td, min_evts);
	108	if (ret > 0)
	109	td->cur_depth -= ret;
40511301	110	}
155f2f02	111
d89ee9f4 BVA	112	if (error \|\| td->error) {
d89ee9f4 BVA	113	pthread_mutex_lock(&td->io_u_lock);
d28174f0	114	pthread_cond_signal(&td->parent->free_cond);
d89ee9f4 BVA	115	pthread_mutex_unlock(&td->io_u_lock);
d89ee9f4 BVA	116	}
d28174f0	117
155f2f02	118	return 0;
40511301 JA	119	}
	120
	121	static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
	122	{
b86ad8f1	123	struct thread_data *td = sw->priv;
40511301	124
d28174f0 JA	125	if (td->error)
d28174f0 JA	126	return false;
40511301 JA	127	if (td->io_u_queued \|\| td->cur_depth \|\| td->io_u_in_flight)
	128	return true;
	129
	130	return false;
	131	}
	132
	133	static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
	134	{
b86ad8f1	135	struct thread_data *td = sw->priv;
40511301 JA	136	int ret;
	137
	138	ret = io_u_quiesce(td);
	139	if (ret > 0)
	140	td->cur_depth -= ret;
	141	}
	142
	143	static int io_workqueue_alloc_fn(struct submit_worker *sw)
	144	{
	145	struct thread_data *td;
	146
	147	td = calloc(1, sizeof(*td));
b86ad8f1	148	sw->priv = td;
40511301 JA	149	return 0;
	150	}
	151
	152	static void io_workqueue_free_fn(struct submit_worker *sw)
	153	{
b86ad8f1 CB	154	free(sw->priv);
b86ad8f1 CB	155	sw->priv = NULL;
40511301 JA	156	}
	157
	158	static int io_workqueue_init_worker_fn(struct submit_worker *sw)
	159	{
	160	struct thread_data *parent = sw->wq->td;
b86ad8f1	161	struct thread_data *td = sw->priv;
40511301 JA	162
	163	memcpy(&td->o, &parent->o, sizeof(td->o));
	164	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	165	td->o.uid = td->o.gid = -1U;
	166	dup_files(td, parent);
	167	td->eo = parent->eo;
	168	fio_options_mem_dupe(td);
	169
	170	if (ioengine_load(td))
	171	goto err;
	172
40511301 JA	173	td->pid = gettid();
	174
	175	INIT_FLIST_HEAD(&td->io_log_list);
	176	INIT_FLIST_HEAD(&td->io_hist_list);
	177	INIT_FLIST_HEAD(&td->verify_list);
	178	INIT_FLIST_HEAD(&td->trim_list);
40511301 JA	179	td->io_hist_tree = RB_ROOT;
	180
	181	td->o.iodepth = 1;
	182	if (td_io_init(td))
	183	goto err_io_init;
	184
4c085cf2 VF	185	if (td->io_ops->post_init && td->io_ops->post_init(td))
	186	goto err_io_init;
	187
d5b3cfd4	188	set_epoch_time(td, td->o.log_unix_epoch \| td->o.log_alternate_epoch, td->o.log_alternate_epoch_clock_id);
40511301 JA	189	fio_getrusage(&td->ru_start);
	190	clear_io_state(td, 1);
	191
	192	td_set_runstate(td, TD_RUNNING);
b7aae4ba	193	td->flags \|= TD_F_CHILD \| TD_F_NEED_LOCK;
40511301 JA	194	td->parent = parent;
	195	return 0;
	196
	197	err_io_init:
	198	close_ioengine(td);
	199	err:
	200	return 1;
	201
	202	}
	203
	204	static void io_workqueue_exit_worker_fn(struct submit_worker *sw,
	205	unsigned int *sum_cnt)
	206	{
b86ad8f1	207	struct thread_data *td = sw->priv;
40511301 JA	208
40511301 JA	209	(*sum_cnt)++;
691310e2 NC	210
	211	/*
	212	* io_workqueue_update_acct_fn() doesn't support per prio stats, and
	213	* even if it did, offload can't be used with all async IO engines.
	214	* If group reporting is set in the parent td, the group result
	215	* generated by __show_run_stats() can still contain multiple prios
	216	* from different offloaded jobs.
	217	*/
	218	sw->wq->td->ts.disable_prio_stat = 1;
016869be	219	sum_thread_stats(&sw->wq->td->ts, &td->ts);
40511301 JA	220
	221	fio_options_free(td);
	222	close_and_free_files(td);
	223	if (td->io_ops)
	224	close_ioengine(td);
	225	td_set_runstate(td, TD_EXITED);
	226	}
	227
	228	#ifdef CONFIG_SFAA
	229	static void sum_val(uint64_t dst, uint64_t src)
	230	{
	231	if (*src) {
	232	__sync_fetch_and_add(dst, *src);
	233	*src = 0;
	234	}
	235	}
	236	#else
	237	static void sum_val(uint64_t dst, uint64_t src)
	238	{
	239	if (*src) {
	240	dst += src;
	241	*src = 0;
	242	}
	243	}
	244	#endif
	245
	246	static void pthread_double_unlock(pthread_mutex_t *lock1,
	247	pthread_mutex_t *lock2)
	248	{
	249	#ifndef CONFIG_SFAA
	250	pthread_mutex_unlock(lock1);
	251	pthread_mutex_unlock(lock2);
	252	#endif
	253	}
	254
	255	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	256	{
	257	#ifndef CONFIG_SFAA
	258	if (lock1 < lock2) {
	259	pthread_mutex_lock(lock1);
	260	pthread_mutex_lock(lock2);
	261	} else {
	262	pthread_mutex_lock(lock2);
	263	pthread_mutex_lock(lock1);
	264	}
	265	#endif
	266	}
	267
	268	static void sum_ddir(struct thread_data dst, struct thread_data src,
	269	enum fio_ddir ddir)
	270	{
	271	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	272
	273	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	274	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	275	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	276	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	277	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
	278
	279	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
	280	}
	281
	282	static void io_workqueue_update_acct_fn(struct submit_worker *sw)
	283	{
b86ad8f1	284	struct thread_data *src = sw->priv;
40511301 JA	285	struct thread_data *dst = sw->wq->td;
	286
	287	if (td_read(src))
	288	sum_ddir(dst, src, DDIR_READ);
	289	if (td_write(src))
	290	sum_ddir(dst, src, DDIR_WRITE);
	291	if (td_trim(src))
	292	sum_ddir(dst, src, DDIR_TRIM);
	293
	294	}
	295
103b174e	296	static struct workqueue_ops rated_wq_ops = {
40511301 JA	297	.fn = io_workqueue_fn,
	298	.pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
	299	.pre_sleep_fn = io_workqueue_pre_sleep_fn,
	300	.update_acct_fn = io_workqueue_update_acct_fn,
	301	.alloc_worker_fn = io_workqueue_alloc_fn,
	302	.free_worker_fn = io_workqueue_free_fn,
	303	.init_worker_fn = io_workqueue_init_worker_fn,
	304	.exit_worker_fn = io_workqueue_exit_worker_fn,
	305	};
103b174e	306
24660963	307	int rate_submit_init(struct thread_data td, struct sk_out sk_out)
103b174e JA	308	{
	309	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	310	return 0;
	311
24660963	312	return workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth, sk_out);
103b174e JA	313	}
	314
	315	void rate_submit_exit(struct thread_data *td)
	316	{
	317	if (td->o.io_submit_mode != IO_MODE_OFFLOAD)
	318	return;
	319
	320	workqueue_exit(&td->io_wq);
	321	}