[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.
 */
int workqueue_enqueue(struct workqueue *wq, struct io_u *io_u)
{
	struct submit_worker *sw;

	sw = get_submit_worker(wq);
	if (sw) {
		const enum fio_ddir ddir = acct_ddir(io_u);
		struct thread_data *parent = wq->td;

		if (ddir_rw(ddir)) {
			parent->io_issues[ddir]++;
			parent->io_issue_bytes[ddir] += io_u->xfer_buflen;
			parent->rate_io_issue_bytes[ddir] += io_u->xfer_buflen;
		}

		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 FIO_Q_QUEUED;
	}

	return FIO_Q_BUSY;
}

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);
	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;
			sw->flags |= SW_F_ACCOUNTED;
			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	/*
	113	* Must be serialized by caller.
	114	*/
	115	int workqueue_enqueue(struct workqueue wq, struct io_u io_u)
	116	{
	117	struct submit_worker *sw;
	118
	119	sw = get_submit_worker(wq);
	120	if (sw) {
	121	const enum fio_ddir ddir = acct_ddir(io_u);
	122	struct thread_data *parent = wq->td;
	123
	124	if (ddir_rw(ddir)) {
	125	parent->io_issues[ddir]++;
	126	parent->io_issue_bytes[ddir] += io_u->xfer_buflen;
50a8ce86	127	parent->rate_io_issue_bytes[ddir] += io_u->xfer_buflen;
a9da8ab2 JA	128	}
	129
	130	pthread_mutex_lock(&sw->lock);
	131	flist_add_tail(&io_u->verify_list, &sw->work_list);
	132	sw->seq = ++wq->work_seq;
	133	sw->flags &= ~SW_F_IDLE;
	134	pthread_mutex_unlock(&sw->lock);
	135
	136	pthread_cond_signal(&sw->cond);
	137	return FIO_Q_QUEUED;
	138	}
	139
	140	return FIO_Q_BUSY;
	141	}
	142
	143	static void handle_list(struct submit_worker sw, struct flist_head list)
	144	{
	145	struct workqueue *wq = sw->wq;
	146	struct io_u *io_u;
	147
	148	while (!flist_empty(list)) {
	149	io_u = flist_first_entry(list, struct io_u, verify_list);
	150	flist_del_init(&io_u->verify_list);
	151	wq->fn(&sw->td, io_u);
	152	}
	153	}
	154
	155	static int init_submit_worker(struct submit_worker *sw)
	156	{
	157	struct thread_data *parent = sw->wq->td;
	158	struct thread_data *td = &sw->td;
	159	int fio_unused ret;
	160
	161	memcpy(&td->o, &parent->o, sizeof(td->o));
	162	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
	163	td->o.uid = td->o.gid = -1U;
	164	dup_files(td, parent);
736a50cd	165	td->eo = parent->eo;
a9da8ab2 JA	166	fio_options_mem_dupe(td);
	167
	168	if (ioengine_load(td))
	169	goto err;
	170
	171	if (td->o.odirect)
	172	td->io_ops->flags \|= FIO_RAWIO;
	173
	174	td->pid = gettid();
	175
	176	INIT_FLIST_HEAD(&td->io_log_list);
	177	INIT_FLIST_HEAD(&td->io_hist_list);
	178	INIT_FLIST_HEAD(&td->verify_list);
	179	INIT_FLIST_HEAD(&td->trim_list);
	180	INIT_FLIST_HEAD(&td->next_rand_list);
	181	td->io_hist_tree = RB_ROOT;
	182
	183	td->o.iodepth = 1;
	184	if (td_io_init(td))
	185	goto err_io_init;
	186
	187	fio_gettime(&td->epoch, NULL);
	188	fio_getrusage(&td->ru_start);
ac28d905	189	clear_io_state(td, 1);
a9da8ab2 JA	190
	191	td_set_runstate(td, TD_RUNNING);
	192	td->flags \|= TD_F_CHILD;
	193	td->parent = parent;
	194	return 0;
	195
	196	err_io_init:
	197	close_ioengine(td);
	198	err:
	199	return 1;
	200	}
	201
f5cd2907	202	#ifdef CONFIG_SFAA
a9da8ab2 JA	203	static void sum_val(uint64_t dst, uint64_t src)
	204	{
	205	if (*src) {
	206	__sync_fetch_and_add(dst, *src);
	207	*src = 0;
	208	}
	209	}
2a274336 JA	210	#else
	211	static void sum_val(uint64_t dst, uint64_t src)
	212	{
	213	if (*src) {
	214	dst += src;
	215	*src = 0;
	216	}
	217	}
	218	#endif
a9da8ab2	219
a6a3469e JA	220	static void pthread_double_unlock(pthread_mutex_t *lock1,
a6a3469e JA	221	pthread_mutex_t *lock2)
a9da8ab2	222	{
f5cd2907	223	#ifndef CONFIG_SFAA
a6a3469e JA	224	pthread_mutex_unlock(lock1);
a6a3469e JA	225	pthread_mutex_unlock(lock2);
2a274336	226	#endif
a6a3469e JA	227	}
	228
	229	static void pthread_double_lock(pthread_mutex_t lock1, pthread_mutex_t lock2)
	230	{
	231	#ifndef CONFIG_SFAA
	232	if (lock1 < lock2) {
	233	pthread_mutex_lock(lock1);
	234	pthread_mutex_lock(lock2);
	235	} else {
	236	pthread_mutex_lock(lock2);
	237	pthread_mutex_lock(lock1);
	238	}
	239	#endif
	240	}
	241
	242	static void sum_ddir(struct thread_data dst, struct thread_data src,
	243	enum fio_ddir ddir)
	244	{
	245	pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
2a274336	246
a9da8ab2 JA	247	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	248	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	249	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	250	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	251	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
2a274336	252
a6a3469e	253	pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
a9da8ab2 JA	254	}
	255
	256	static void update_accounting(struct submit_worker *sw)
	257	{
	258	struct thread_data *src = &sw->td;
	259	struct thread_data *dst = sw->wq->td;
	260
	261	if (td_read(src))
	262	sum_ddir(dst, src, DDIR_READ);
	263	if (td_write(src))
	264	sum_ddir(dst, src, DDIR_WRITE);
	265	if (td_trim(src))
	266	sum_ddir(dst, src, DDIR_TRIM);
	267	}
	268
	269	static void worker_thread(void data)
	270	{
	271	struct submit_worker *sw = data;
	272	struct workqueue *wq = sw->wq;
	273	struct thread_data *td = &sw->td;
	274	unsigned int eflags = 0, ret;
	275	FLIST_HEAD(local_list);
	276
	277	ret = init_submit_worker(sw);
	278	pthread_mutex_lock(&sw->lock);
	279	sw->flags \|= SW_F_RUNNING;
	280	if (ret)
	281	sw->flags \|= SW_F_ERROR;
	282	pthread_mutex_unlock(&sw->lock);
	283
	284	pthread_mutex_lock(&wq->flush_lock);
	285	pthread_cond_signal(&wq->flush_cond);
	286	pthread_mutex_unlock(&wq->flush_lock);
	287
	288	if (sw->flags & SW_F_ERROR)
	289	goto done;
	290
	291	while (1) {
	292	pthread_mutex_lock(&sw->lock);
	293
	294	if (flist_empty(&sw->work_list)) {
	295	if (sw->flags & SW_F_EXIT) {
	296	pthread_mutex_unlock(&sw->lock);
	297	break;
	298	}
	299
	300	if (td->io_u_queued \|\| td->cur_depth \|\|
	301	td->io_u_in_flight) {
40649b00 JA	302	int ret;
40649b00 JA	303
a9da8ab2	304	pthread_mutex_unlock(&sw->lock);
40649b00 JA	305	ret = io_u_quiesce(td);
	306	if (ret > 0)
	307	td->cur_depth -= ret;
a9da8ab2 JA	308	pthread_mutex_lock(&sw->lock);
	309	}
	310
	311	/*
	312	* We dropped and reaquired the lock, check
	313	* state again.
	314	*/
	315	if (!flist_empty(&sw->work_list))
	316	goto handle_work;
	317
	318	if (sw->flags & SW_F_EXIT) {
	319	pthread_mutex_unlock(&sw->lock);
	320	break;
	321	} else if (!(sw->flags & SW_F_IDLE)) {
	322	sw->flags \|= SW_F_IDLE;
	323	wq->next_free_worker = sw->index;
	324	if (wq->wake_idle)
	325	pthread_cond_signal(&wq->flush_cond);
	326	}
	327	update_accounting(sw);
	328	pthread_cond_wait(&sw->cond, &sw->lock);
	329	} else {
	330	handle_work:
	331	flist_splice_init(&sw->work_list, &local_list);
	332	}
	333	pthread_mutex_unlock(&sw->lock);
	334	handle_list(sw, &local_list);
	335	}
	336
	337	update_accounting(sw);
	338
	339	done:
	340	pthread_mutex_lock(&sw->lock);
	341	sw->flags \|= (SW_F_EXITED \| eflags);
	342	pthread_mutex_unlock(&sw->lock);
	343	return NULL;
	344	}
	345
	346	static void free_worker(struct submit_worker *sw)
	347	{
	348	struct thread_data *td = &sw->td;
	349
	350	fio_options_free(td);
	351	close_and_free_files(td);
	352	if (td->io_ops)
	353	close_ioengine(td);
	354	td_set_runstate(td, TD_EXITED);
	355
	356	pthread_cond_destroy(&sw->cond);
	357	pthread_mutex_destroy(&sw->lock);
	358	}
	359
	360	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	361	{
	362	struct thread_data *parent = sw->wq->td;
	363
	364	pthread_join(sw->thread, NULL);
	365	(*sum_cnt)++;
	366	sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt);
	367	free_worker(sw);
	368	}
	369
	370	void workqueue_exit(struct workqueue *wq)
	371	{
372	unsigned int shutdown, sum_cnt = 0;
373	struct submit_worker *sw;
374	int i;
375
376	for (i = 0; i < wq->max_workers; i++) {
377	sw = &wq->workers[i];
378
379	pthread_mutex_lock(&sw->lock);
380	sw->flags \|= SW_F_EXIT;
381	pthread_cond_signal(&sw->cond);
382	pthread_mutex_unlock(&sw->lock);
383	}
384
385	do {
386	shutdown = 0;
387	for (i = 0; i < wq->max_workers; i++) {
388	sw = &wq->workers[i];
389	if (sw->flags & SW_F_ACCOUNTED)
390	continue;
391	sw->flags \|= SW_F_ACCOUNTED;
392	shutdown_worker(sw, &sum_cnt);
393	shutdown++;
394	}
395	} while (shutdown && shutdown != wq->max_workers);
396
397	free(wq->workers);
398	pthread_mutex_destroy(&wq->flush_lock);
399	pthread_cond_destroy(&wq->flush_cond);
2a274336	400	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	401	}
	402
	403	static int start_worker(struct workqueue *wq, unsigned int index)
	404	{
	405	struct submit_worker *sw = &wq->workers[index];
	406	int ret;
	407
	408	INIT_FLIST_HEAD(&sw->work_list);
	409	pthread_cond_init(&sw->cond, NULL);
	410	pthread_mutex_init(&sw->lock, NULL);
	411	sw->wq = wq;
	412	sw->index = index;
	413
	414	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	415	if (!ret) {
	416	pthread_mutex_lock(&sw->lock);
	417	sw->flags = SW_F_IDLE;
	418	pthread_mutex_unlock(&sw->lock);
	419	return 0;
	420	}
	421
	422	free_worker(sw);
	423	return 1;
	424	}
	425
	426	int workqueue_init(struct thread_data td, struct workqueue wq,
	427	workqueue_fn *fn, unsigned max_pending)
	428	{
	429	unsigned int running;
	430	int i, error;
	431
	432	wq->max_workers = max_pending;
	433	wq->td = td;
	434	wq->fn = fn;
	435	wq->work_seq = 0;
	436	wq->next_free_worker = 0;
	437	pthread_cond_init(&wq->flush_cond, NULL);
	438	pthread_mutex_init(&wq->flush_lock, NULL);
2a274336	439	pthread_mutex_init(&wq->stat_lock, NULL);
a9da8ab2 JA	440
	441	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
	442
	443	for (i = 0; i < wq->max_workers; i++)
	444	if (start_worker(wq, i))
	445	break;
	446
	447	wq->max_workers = i;
bddc8d16 JA	448	if (!wq->max_workers)
bddc8d16 JA	449	goto err;
a9da8ab2 JA	450
	451	/*
	452	* Wait for them all to be started and initialized
	453	*/
	454	error = 0;
	455	do {
	456	struct submit_worker *sw;
	457
	458	running = 0;
	459	pthread_mutex_lock(&wq->flush_lock);
	460	for (i = 0; i < wq->max_workers; i++) {
	461	sw = &wq->workers[i];
	462	pthread_mutex_lock(&sw->lock);
	463	if (sw->flags & SW_F_RUNNING)
	464	running++;
	465	if (sw->flags & SW_F_ERROR)
	466	error++;
	467	pthread_mutex_unlock(&sw->lock);
	468	}
	469
	470	if (error \|\| running == wq->max_workers) {
	471	pthread_mutex_unlock(&wq->flush_lock);
	472	break;
	473	}
	474
	475	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	476	pthread_mutex_unlock(&wq->flush_lock);
	477	} while (1);
	478
bddc8d16 JA	479	if (!error)
bddc8d16 JA	480	return 0;
a9da8ab2	481
bddc8d16 JA	482	err:
	483	log_err("Can't create rate workqueue\n");
	484	td_verror(td, ESRCH, "workqueue_init");
	485	workqueue_exit(wq);
	486	return 1;
a9da8ab2	487	}