[fio.git] / workqueue.c

/*
 * Rated submission helpers
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include <unistd.h>

#include "fio.h"
#include "ioengine.h"
#include "flist.h"
#include "workqueue.h"
#include "lib/getrusage.h"

struct submit_worker {
	pthread_t thread;
	pthread_mutex_t lock;
	pthread_cond_t cond;
	struct flist_head work_list;
	unsigned int flags;
	unsigned int index;
	uint64_t seq;
	struct workqueue *wq;
	struct thread_data td;
};

enum {
	SW_F_IDLE	= 1 << 0,
	SW_F_RUNNING	= 1 << 1,
	SW_F_EXIT	= 1 << 2,
	SW_F_EXITED	= 1 << 3,
	SW_F_ACCOUNTED	= 1 << 4,
	SW_F_ERROR	= 1 << 5,
};

static struct submit_worker *__get_submit_worker(struct workqueue *wq,
						 unsigned int start,
						 unsigned int end,
						 struct submit_worker **best)
{
	struct submit_worker *sw = NULL;

	while (start <= end) {
		sw = &wq->workers[start];
		if (sw->flags & SW_F_IDLE)
			return sw;
		if (!(*best) || sw->seq < (*best)->seq)
			*best = sw;
		start++;
	}

	return NULL;
}

static struct submit_worker *get_submit_worker(struct workqueue *wq)
{
	unsigned int next = wq->next_free_worker;
	struct submit_worker *sw, *best = NULL;

	assert(next < wq->max_workers);

	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	if (!sw && next)
		sw = __get_submit_worker(wq, 0, next - 1, &best);

	/*
	 * No truly idle found, use best match
	 */
	if (!sw)
		sw = best;

	if (sw->index == wq->next_free_worker) {
		if (sw->index + 1 < wq->max_workers)
			wq->next_free_worker = sw->index + 1;
		else
			wq->next_free_worker = 0;
	}

	return sw;
}

static bool all_sw_idle(struct workqueue *wq)
{
	int i;

	for (i = 0; i < wq->max_workers; i++) {
		struct submit_worker *sw = &wq->workers[i];

		if (!(sw->flags & SW_F_IDLE))
			return false;
	}

	return true;
}

/*
 * Must be serialized wrt workqueue_enqueue() by caller
 */
void workqueue_flush(struct workqueue *wq)
{
	wq->wake_idle = 1;

	while (!all_sw_idle(wq)) {
		pthread_mutex_lock(&wq->flush_lock);
		pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
		pthread_mutex_unlock(&wq->flush_lock);
	}

	wq->wake_idle = 0;
}

/*
 * Must be serialized by caller. Returns true for queued, false for busy.
 */
bool workqueue_enqueue(struct workqueue *wq, struct io_u *io_u)
{
	struct submit_worker *sw;

	sw = get_submit_worker(wq);
	if (sw) {
		pthread_mutex_lock(&sw->lock);
		flist_add_tail(&io_u->verify_list, &sw->work_list);
		sw->seq = ++wq->work_seq;
		sw->flags &= ~SW_F_IDLE;
		pthread_mutex_unlock(&sw->lock);

		pthread_cond_signal(&sw->cond);
		return true;
	}

	return false;
}

static void handle_list(struct submit_worker *sw, struct flist_head *list)
{
	struct workqueue *wq = sw->wq;
	struct io_u *io_u;

	while (!flist_empty(list)) {
		io_u = flist_first_entry(list, struct io_u, verify_list);
		flist_del_init(&io_u->verify_list);
		wq->fn(&sw->td, io_u);
	}
}

static int init_submit_worker(struct submit_worker *sw)
{
	struct thread_data *parent = sw->wq->td;
	struct thread_data *td = &sw->td;
	int fio_unused ret;

	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;

	if (td->o.odirect)
		td->io_ops->flags |= FIO_RAWIO;

	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);
	INIT_FLIST_HEAD(&td->next_rand_list);
	td->io_hist_tree = RB_ROOT;

	td->o.iodepth = 1;
	if (td_io_init(td))
		goto err_io_init;

	fio_gettime(&td->epoch, NULL);
	fio_getrusage(&td->ru_start);
	clear_io_state(td, 1);

	td_set_runstate(td, TD_RUNNING);
	td->flags |= TD_F_CHILD;
	td->parent = parent;
	return 0;

err_io_init:
	close_ioengine(td);
err:
	return 1;
}

#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 update_accounting(struct submit_worker *sw)
{
	struct thread_data *src = &sw->td;
	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 void *worker_thread(void *data)
{
	struct submit_worker *sw = data;
	struct workqueue *wq = sw->wq;
	struct thread_data *td = &sw->td;
	unsigned int eflags = 0, ret;
	FLIST_HEAD(local_list);

	ret = init_submit_worker(sw);
	pthread_mutex_lock(&sw->lock);
	sw->flags |= SW_F_RUNNING;
	if (ret)
		sw->flags |= SW_F_ERROR;
	pthread_mutex_unlock(&sw->lock);

	pthread_mutex_lock(&wq->flush_lock);
	pthread_cond_signal(&wq->flush_cond);
	pthread_mutex_unlock(&wq->flush_lock);

	if (sw->flags & SW_F_ERROR)
		goto done;

	while (1) {
		pthread_mutex_lock(&sw->lock);

		if (flist_empty(&sw->work_list)) {
			if (sw->flags & SW_F_EXIT) {
				pthread_mutex_unlock(&sw->lock);
				break;
			}

			if (td->io_u_queued || td->cur_depth ||
			    td->io_u_in_flight) {
				int ret;

				pthread_mutex_unlock(&sw->lock);
				ret = io_u_quiesce(td);
				if (ret > 0)
					td->cur_depth -= ret;
				pthread_mutex_lock(&sw->lock);
			}

			/*
			 * We dropped and reaquired the lock, check
			 * state again.
			 */
			if (!flist_empty(&sw->work_list))
				goto handle_work;

			if (sw->flags & SW_F_EXIT) {
				pthread_mutex_unlock(&sw->lock);
				break;
			} else if (!(sw->flags & SW_F_IDLE)) {
				sw->flags |= SW_F_IDLE;
				wq->next_free_worker = sw->index;
				if (wq->wake_idle)
					pthread_cond_signal(&wq->flush_cond);
			}
			update_accounting(sw);
			pthread_cond_wait(&sw->cond, &sw->lock);
		} else {
handle_work:
			flist_splice_init(&sw->work_list, &local_list);
		}
		pthread_mutex_unlock(&sw->lock);
		handle_list(sw, &local_list);
	}

	update_accounting(sw);

done:
	pthread_mutex_lock(&sw->lock);
	sw->flags |= (SW_F_EXITED | eflags);
	pthread_mutex_unlock(&sw->lock);
	return NULL;
}

static void free_worker(struct submit_worker *sw)
{
	struct thread_data *td = &sw->td;

	fio_options_free(td);
	close_and_free_files(td);
	if (td->io_ops)
		close_ioengine(td);
	td_set_runstate(td, TD_EXITED);

	pthread_cond_destroy(&sw->cond);
	pthread_mutex_destroy(&sw->lock);
}

static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
{
	struct thread_data *parent = sw->wq->td;

	pthread_join(sw->thread, NULL);
	(*sum_cnt)++;
	sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt == 1);
	free_worker(sw);
}

void workqueue_exit(struct workqueue *wq)
{
	unsigned int shutdown, sum_cnt = 0;
	struct submit_worker *sw;
	int i;

	for (i = 0; i < wq->max_workers; i++) {
		sw = &wq->workers[i];

		pthread_mutex_lock(&sw->lock);
		sw->flags |= SW_F_EXIT;
		pthread_cond_signal(&sw->cond);
		pthread_mutex_unlock(&sw->lock);
	}

	do {
		shutdown = 0;
		for (i = 0; i < wq->max_workers; i++) {
			sw = &wq->workers[i];
			if (sw->flags & SW_F_ACCOUNTED)
				continue;
			pthread_mutex_lock(&sw->lock);
			sw->flags |= SW_F_ACCOUNTED;
			pthread_mutex_unlock(&sw->lock);
			shutdown_worker(sw, &sum_cnt);
			shutdown++;
		}
	} while (shutdown && shutdown != wq->max_workers);

	free(wq->workers);
	pthread_mutex_destroy(&wq->flush_lock);
	pthread_cond_destroy(&wq->flush_cond);
	pthread_mutex_destroy(&wq->stat_lock);
}

static int start_worker(struct workqueue *wq, unsigned int index)
{
	struct submit_worker *sw = &wq->workers[index];
	int ret;

	INIT_FLIST_HEAD(&sw->work_list);
	pthread_cond_init(&sw->cond, NULL);
	pthread_mutex_init(&sw->lock, NULL);
	sw->wq = wq;
	sw->index = index;

	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	if (!ret) {
		pthread_mutex_lock(&sw->lock);
		sw->flags = SW_F_IDLE;
		pthread_mutex_unlock(&sw->lock);
		return 0;
	}

	free_worker(sw);
	return 1;
}

int workqueue_init(struct thread_data *td, struct workqueue *wq,
		   workqueue_fn *fn, unsigned max_pending)
{
	unsigned int running;
	int i, error;

	wq->max_workers = max_pending;
	wq->td = td;
	wq->fn = fn;
	wq->work_seq = 0;
	wq->next_free_worker = 0;
	pthread_cond_init(&wq->flush_cond, NULL);
	pthread_mutex_init(&wq->flush_lock, NULL);
	pthread_mutex_init(&wq->stat_lock, NULL);

	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));

	for (i = 0; i < wq->max_workers; i++)
		if (start_worker(wq, i))
			break;

	wq->max_workers = i;
	if (!wq->max_workers)
		goto err;

	/*
	 * Wait for them all to be started and initialized
	 */
	error = 0;
	do {
		struct submit_worker *sw;

		running = 0;
		pthread_mutex_lock(&wq->flush_lock);
		for (i = 0; i < wq->max_workers; i++) {
			sw = &wq->workers[i];
			pthread_mutex_lock(&sw->lock);
			if (sw->flags & SW_F_RUNNING)
				running++;
			if (sw->flags & SW_F_ERROR)
				error++;
			pthread_mutex_unlock(&sw->lock);
		}

		if (error || running == wq->max_workers) {
			pthread_mutex_unlock(&wq->flush_lock);
			break;
		}

		pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
		pthread_mutex_unlock(&wq->flush_lock);
	} while (1);

	if (!error)
		return 0;

err:
	log_err("Can't create rate workqueue\n");
	td_verror(td, ESRCH, "workqueue_init");
	workqueue_exit(wq);
	return 1;
}
Commit	Line	Data
a9da8ab2 JA	1	/*
	2	* Rated submission helpers
	3	*
	4	* Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
	5	*
	6	*/
	7	#include <unistd.h>
	8
	9	#include "fio.h"
	10	#include "ioengine.h"
	11	#include "flist.h"
	12	#include "workqueue.h"
	13	#include "lib/getrusage.h"
	14
	15	struct submit_worker {
	16	pthread_t thread;
	17	pthread_mutex_t lock;
	18	pthread_cond_t cond;
	19	struct flist_head work_list;
	20	unsigned int flags;
	21	unsigned int index;
	22	uint64_t seq;
	23	struct workqueue *wq;
	24	struct thread_data td;
	25	};
	26
	27	enum {
	28	SW_F_IDLE = 1 << 0,
	29	SW_F_RUNNING = 1 << 1,
	30	SW_F_EXIT = 1 << 2,
	31	SW_F_EXITED = 1 << 3,
	32	SW_F_ACCOUNTED = 1 << 4,
	33	SW_F_ERROR = 1 << 5,
	34	};
	35
	36	static struct submit_worker __get_submit_worker(struct workqueue wq,
	37	unsigned int start,
	38	unsigned int end,
	39	struct submit_worker **best)
	40	{
	41	struct submit_worker *sw = NULL;
	42
	43	while (start <= end) {
	44	sw = &wq->workers[start];
	45	if (sw->flags & SW_F_IDLE)
	46	return sw;
	47	if (!(best) \|\| sw->seq < (best)->seq)
	48	*best = sw;
	49	start++;
	50	}
	51
	52	return NULL;
	53	}
	54
	55	static struct submit_worker get_submit_worker(struct workqueue wq)
	56	{
	57	unsigned int next = wq->next_free_worker;
	58	struct submit_worker sw, best = NULL;
	59
	60	assert(next < wq->max_workers);
	61
	62	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	63	if (!sw && next)
	64	sw = __get_submit_worker(wq, 0, next - 1, &best);
65
66	/*
67	* No truly idle found, use best match
68	*/
69	if (!sw)
70	sw = best;
71
72	if (sw->index == wq->next_free_worker) {
73	if (sw->index + 1 < wq->max_workers)
74	wq->next_free_worker = sw->index + 1;
75	else
76	wq->next_free_worker = 0;
77	}
78
79	return sw;
80	}
81
1391052a	82	static bool all_sw_idle(struct workqueue *wq)
a9da8ab2 JA	83	{
	84	int i;
	85
	86	for (i = 0; i < wq->max_workers; i++) {
	87	struct submit_worker *sw = &wq->workers[i];
	88
	89	if (!(sw->flags & SW_F_IDLE))
1391052a	90	return false;
a9da8ab2 JA	91	}
a9da8ab2 JA	92
1391052a	93	return true;
a9da8ab2 JA	94	}
	95
	96	/*
	97	* Must be serialized wrt workqueue_enqueue() by caller
	98	*/
	99	void workqueue_flush(struct workqueue *wq)
	100	{
	101	wq->wake_idle = 1;
	102
	103	while (!all_sw_idle(wq)) {
	104	pthread_mutex_lock(&wq->flush_lock);
	105	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	106	pthread_mutex_unlock(&wq->flush_lock);
	107	}
	108
	109	wq->wake_idle = 0;
	110	}
	111
	112	/*
b07f6ad1	113	* Must be serialized by caller. Returns true for queued, false for busy.
a9da8ab2	114	*/
b07f6ad1	115	bool workqueue_enqueue(struct workqueue wq, struct io_u io_u)
a9da8ab2 JA	116	{
	117	struct submit_worker *sw;
	118
	119	sw = get_submit_worker(wq);
	120	if (sw) {
a9da8ab2 JA	121	pthread_mutex_lock(&sw->lock);
	122	flist_add_tail(&io_u->verify_list, &sw->work_list);
	123	sw->seq = ++wq->work_seq;
	124	sw->flags &= ~SW_F_IDLE;
	125	pthread_mutex_unlock(&sw->lock);
	126
	127	pthread_cond_signal(&sw->cond);
b07f6ad1	128	return true;
a9da8ab2 JA	129	}
a9da8ab2 JA	130
b07f6ad1	131	return false;
a9da8ab2 JA	132	}
	133
	134	static void handle_list(struct submit_worker sw, struct flist_head list)
	135	{
	136	struct workqueue *wq = sw->wq;
	137	struct io_u *io_u;
	138
	139	while (!flist_empty(list)) {
	140	io_u = flist_first_entry(list, struct io_u, verify_list);
	141	flist_del_init(&io_u->verify_list);
	142	wq->fn(&sw->td, io_u);
	143	}
	144	}
	145
	146	static int init_submit_worker(struct submit_worker *sw)
	147	{
	148	struct thread_data *parent = sw->wq->td;
	149	struct thread_data *td = &sw->td;
	150	int fio_unused ret;
	151
	152	memcpy(&td->o, &parent->o, sizeof(td->o));
	153	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	154	td->o.uid = td->o.gid = -1U;
	155	dup_files(td, parent);
736a50cd	156	td->eo = parent->eo;
a9da8ab2 JA	157	fio_options_mem_dupe(td);
	158
	159	if (ioengine_load(td))
	160	goto err;
	161
	162	if (td->o.odirect)
	163	td->io_ops->flags \|= FIO_RAWIO;
	164
	165	td->pid = gettid();
	166
	167	INIT_FLIST_HEAD(&td->io_log_list);
	168	INIT_FLIST_HEAD(&td->io_hist_list);
	169	INIT_FLIST_HEAD(&td->verify_list);
	170	INIT_FLIST_HEAD(&td->trim_list);
	171	INIT_FLIST_HEAD(&td->next_rand_list);
	172	td->io_hist_tree = RB_ROOT;
	173
	174	td->o.iodepth = 1;
	175	if (td_io_init(td))
	176	goto err_io_init;
	177
	178	fio_gettime(&td->epoch, NULL);
	179	fio_getrusage(&td->ru_start);
ac28d905	180	clear_io_state(td, 1);
a9da8ab2 JA	181
	182	td_set_runstate(td, TD_RUNNING);
	183	td->flags \|= TD_F_CHILD;
	184	td->parent = parent;
	185	return 0;
	186
	187	err_io_init:
	188	close_ioengine(td);
	189	err:
	190	return 1;
	191	}
	192
f5cd2907	193	#ifdef CONFIG_SFAA
a9da8ab2 JA	194	static void sum_val(uint64_t dst, uint64_t src)
	195	{
	196	if (*src) {
	197	__sync_fetch_and_add(dst, *src);
	198	*src = 0;
	199	}
	200	}
2a274336 JA	201	#else
	202	static void sum_val(uint64_t dst, uint64_t src)
	203	{
	204	if (*src) {
	205	dst += src;
	206	*src = 0;
	207	}
	208	}
	209	#endif
a9da8ab2	210
a6a3469e JA	211	static void pthread_double_unlock(pthread_mutex_t *lock1,
a6a3469e JA	212	pthread_mutex_t *lock2)
a9da8ab2	213	{
f5cd2907	214	#ifndef CONFIG_SFAA
a6a3469e JA	215	pthread_mutex_unlock(lock1);
a6a3469e JA	216	pthread_mutex_unlock(lock2);
2a274336	217	#endif
a6a3469e JA	218	}
	219
	220	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	221	{
	222	#ifndef CONFIG_SFAA
	223	if (lock1 < lock2) {
	224	pthread_mutex_lock(lock1);
	225	pthread_mutex_lock(lock2);
	226	} else {
	227	pthread_mutex_lock(lock2);
	228	pthread_mutex_lock(lock1);
	229	}
	230	#endif
	231	}
	232
	233	static void sum_ddir(struct thread_data dst, struct thread_data src,
	234	enum fio_ddir ddir)
	235	{
	236	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
2a274336	237
a9da8ab2 JA	238	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	239	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	240	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	241	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	242	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
2a274336	243
a6a3469e	244	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
a9da8ab2 JA	245	}
	246
	247	static void update_accounting(struct submit_worker *sw)
	248	{
	249	struct thread_data *src = &sw->td;
	250	struct thread_data *dst = sw->wq->td;
	251
	252	if (td_read(src))
	253	sum_ddir(dst, src, DDIR_READ);
	254	if (td_write(src))
	255	sum_ddir(dst, src, DDIR_WRITE);
	256	if (td_trim(src))
	257	sum_ddir(dst, src, DDIR_TRIM);
	258	}
	259
	260	static void worker_thread(void data)
	261	{
	262	struct submit_worker *sw = data;
	263	struct workqueue *wq = sw->wq;
	264	struct thread_data *td = &sw->td;
	265	unsigned int eflags = 0, ret;
	266	FLIST_HEAD(local_list);
	267
	268	ret = init_submit_worker(sw);
	269	pthread_mutex_lock(&sw->lock);
	270	sw->flags \|= SW_F_RUNNING;
	271	if (ret)
	272	sw->flags \|= SW_F_ERROR;
	273	pthread_mutex_unlock(&sw->lock);
	274
	275	pthread_mutex_lock(&wq->flush_lock);
	276	pthread_cond_signal(&wq->flush_cond);
	277	pthread_mutex_unlock(&wq->flush_lock);
	278
	279	if (sw->flags & SW_F_ERROR)
	280	goto done;
	281
	282	while (1) {
	283	pthread_mutex_lock(&sw->lock);
	284
	285	if (flist_empty(&sw->work_list)) {
	286	if (sw->flags & SW_F_EXIT) {
	287	pthread_mutex_unlock(&sw->lock);
	288	break;
	289	}
	290
	291	if (td->io_u_queued \|\| td->cur_depth \|\|
	292	td->io_u_in_flight) {
40649b00 JA	293	int ret;
40649b00 JA	294
a9da8ab2	295	pthread_mutex_unlock(&sw->lock);
40649b00 JA	296	ret = io_u_quiesce(td);
	297	if (ret > 0)
	298	td->cur_depth -= ret;
a9da8ab2 JA	299	pthread_mutex_lock(&sw->lock);
	300	}
	301
	302	/*
	303	* We dropped and reaquired the lock, check
	304	* state again.
	305	*/
	306	if (!flist_empty(&sw->work_list))
	307	goto handle_work;
	308
	309	if (sw->flags & SW_F_EXIT) {
	310	pthread_mutex_unlock(&sw->lock);
	311	break;
	312	} else if (!(sw->flags & SW_F_IDLE)) {
	313	sw->flags \|= SW_F_IDLE;
	314	wq->next_free_worker = sw->index;
	315	if (wq->wake_idle)
	316	pthread_cond_signal(&wq->flush_cond);
	317	}
	318	update_accounting(sw);
	319	pthread_cond_wait(&sw->cond, &sw->lock);
	320	} else {
	321	handle_work:
	322	flist_splice_init(&sw->work_list, &local_list);
	323	}
	324	pthread_mutex_unlock(&sw->lock);
	325	handle_list(sw, &local_list);
	326	}
	327
	328	update_accounting(sw);
	329
	330	done:
	331	pthread_mutex_lock(&sw->lock);
	332	sw->flags \|= (SW_F_EXITED \| eflags);
	333	pthread_mutex_unlock(&sw->lock);
	334	return NULL;
	335	}
	336
	337	static void free_worker(struct submit_worker *sw)
	338	{
	339	struct thread_data *td = &sw->td;
	340
	341	fio_options_free(td);
	342	close_and_free_files(td);
	343	if (td->io_ops)
	344	close_ioengine(td);
	345	td_set_runstate(td, TD_EXITED);
	346
	347	pthread_cond_destroy(&sw->cond);
	348	pthread_mutex_destroy(&sw->lock);
	349	}
	350
	351	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	352	{
	353	struct thread_data *parent = sw->wq->td;
	354
	355	pthread_join(sw->thread, NULL);
	356	(*sum_cnt)++;
fd595830	357	sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt == 1);
a9da8ab2 JA	358	free_worker(sw);
	359	}
	360
	361	void workqueue_exit(struct workqueue *wq)
	362	{
	363	unsigned int shutdown, sum_cnt = 0;
	364	struct submit_worker *sw;
	365	int i;
	366
	367	for (i = 0; i < wq->max_workers; i++) {
	368	sw = &wq->workers[i];
	369
	370	pthread_mutex_lock(&sw->lock);
	371	sw->flags \|= SW_F_EXIT;
	372	pthread_cond_signal(&sw->cond);
	373	pthread_mutex_unlock(&sw->lock);
	374	}
	375
	376	do {
	377	shutdown = 0;
	378	for (i = 0; i < wq->max_workers; i++) {
	379	sw = &wq->workers[i];
	380	if (sw->flags & SW_F_ACCOUNTED)
	381	continue;
b7d0bbf1	382	pthread_mutex_lock(&sw->lock);
a9da8ab2	383	sw->flags \|= SW_F_ACCOUNTED;
b7d0bbf1	384	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	385	shutdown_worker(sw, &sum_cnt);
	386	shutdown++;
	387	}
	388	} while (shutdown && shutdown != wq->max_workers);
	389
	390	free(wq->workers);
	391	pthread_mutex_destroy(&wq->flush_lock);
	392	pthread_cond_destroy(&wq->flush_cond);
2a274336	393	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	394	}
	395
	396	static int start_worker(struct workqueue *wq, unsigned int index)
	397	{
	398	struct submit_worker *sw = &wq->workers[index];
	399	int ret;
	400
	401	INIT_FLIST_HEAD(&sw->work_list);
	402	pthread_cond_init(&sw->cond, NULL);
	403	pthread_mutex_init(&sw->lock, NULL);
	404	sw->wq = wq;
	405	sw->index = index;
	406
	407	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	408	if (!ret) {
	409	pthread_mutex_lock(&sw->lock);
	410	sw->flags = SW_F_IDLE;
	411	pthread_mutex_unlock(&sw->lock);
	412	return 0;
	413	}
	414
	415	free_worker(sw);
	416	return 1;
	417	}
	418
	419	int workqueue_init(struct thread_data td, struct workqueue wq,
	420	workqueue_fn *fn, unsigned max_pending)
	421	{
	422	unsigned int running;
	423	int i, error;
	424
	425	wq->max_workers = max_pending;
	426	wq->td = td;
	427	wq->fn = fn;
	428	wq->work_seq = 0;
	429	wq->next_free_worker = 0;
	430	pthread_cond_init(&wq->flush_cond, NULL);
	431	pthread_mutex_init(&wq->flush_lock, NULL);
2a274336	432	pthread_mutex_init(&wq->stat_lock, NULL);
a9da8ab2 JA	433
	434	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
	435
	436	for (i = 0; i < wq->max_workers; i++)
	437	if (start_worker(wq, i))
	438	break;
	439
	440	wq->max_workers = i;
bddc8d16 JA	441	if (!wq->max_workers)
bddc8d16 JA	442	goto err;
a9da8ab2 JA	443
	444	/*
	445	* Wait for them all to be started and initialized
	446	*/
	447	error = 0;
	448	do {
	449	struct submit_worker *sw;
	450
	451	running = 0;
	452	pthread_mutex_lock(&wq->flush_lock);
	453	for (i = 0; i < wq->max_workers; i++) {
	454	sw = &wq->workers[i];
	455	pthread_mutex_lock(&sw->lock);
	456	if (sw->flags & SW_F_RUNNING)
	457	running++;
	458	if (sw->flags & SW_F_ERROR)
	459	error++;
	460	pthread_mutex_unlock(&sw->lock);
	461	}
	462
	463	if (error \|\| running == wq->max_workers) {
	464	pthread_mutex_unlock(&wq->flush_lock);
	465	break;
	466	}
	467
	468	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	469	pthread_mutex_unlock(&wq->flush_lock);
	470	} while (1);
	471
bddc8d16 JA	472	if (!error)
bddc8d16 JA	473	return 0;
a9da8ab2	474
bddc8d16 JA	475	err:
	476	log_err("Can't create rate workqueue\n");
	477	td_verror(td, ESRCH, "workqueue_init");
	478	workqueue_exit(wq);
	479	return 1;
a9da8ab2	480	}