[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 sum_ddir(struct thread_data *dst, struct thread_data *src,
		     enum fio_ddir ddir)
{
#ifndef CONFIG_SFAA
	pthread_mutex_lock(&dst->io_wq.stat_lock);
	pthread_mutex_lock(&src->io_wq.stat_lock);
#endif

	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]);

#ifndef CONFIG_SFAA
	pthread_mutex_unlock(&src->io_wq.stat_lock);
	pthread_mutex_unlock(&dst->io_wq.stat_lock);
#endif
}

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 JA	219
	220	static void sum_ddir(struct thread_data dst, struct thread_data src,
	221	enum fio_ddir ddir)
	222	{
f5cd2907	223	#ifndef CONFIG_SFAA
2a274336 JA	224	pthread_mutex_lock(&dst->io_wq.stat_lock);
	225	pthread_mutex_lock(&src->io_wq.stat_lock);
	226	#endif
	227
a9da8ab2 JA	228	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
	229	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
	230	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
	231	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
	232	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
2a274336	233
f5cd2907	234	#ifndef CONFIG_SFAA
2a274336 JA	235	pthread_mutex_unlock(&src->io_wq.stat_lock);
	236	pthread_mutex_unlock(&dst->io_wq.stat_lock);
	237	#endif
a9da8ab2 JA	238	}
	239
	240	static void update_accounting(struct submit_worker *sw)
	241	{
	242	struct thread_data *src = &sw->td;
	243	struct thread_data *dst = sw->wq->td;
	244
	245	if (td_read(src))
	246	sum_ddir(dst, src, DDIR_READ);
	247	if (td_write(src))
	248	sum_ddir(dst, src, DDIR_WRITE);
	249	if (td_trim(src))
	250	sum_ddir(dst, src, DDIR_TRIM);
	251	}
	252
	253	static void worker_thread(void data)
	254	{
	255	struct submit_worker *sw = data;
	256	struct workqueue *wq = sw->wq;
	257	struct thread_data *td = &sw->td;
	258	unsigned int eflags = 0, ret;
	259	FLIST_HEAD(local_list);
	260
	261	ret = init_submit_worker(sw);
	262	pthread_mutex_lock(&sw->lock);
	263	sw->flags \|= SW_F_RUNNING;
	264	if (ret)
	265	sw->flags \|= SW_F_ERROR;
	266	pthread_mutex_unlock(&sw->lock);
	267
	268	pthread_mutex_lock(&wq->flush_lock);
	269	pthread_cond_signal(&wq->flush_cond);
	270	pthread_mutex_unlock(&wq->flush_lock);
	271
	272	if (sw->flags & SW_F_ERROR)
	273	goto done;
	274
	275	while (1) {
	276	pthread_mutex_lock(&sw->lock);
	277
	278	if (flist_empty(&sw->work_list)) {
	279	if (sw->flags & SW_F_EXIT) {
	280	pthread_mutex_unlock(&sw->lock);
	281	break;
	282	}
	283
	284	if (td->io_u_queued \|\| td->cur_depth \|\|
	285	td->io_u_in_flight) {
40649b00 JA	286	int ret;
40649b00 JA	287
a9da8ab2	288	pthread_mutex_unlock(&sw->lock);
40649b00 JA	289	ret = io_u_quiesce(td);
	290	if (ret > 0)
	291	td->cur_depth -= ret;
a9da8ab2 JA	292	pthread_mutex_lock(&sw->lock);
	293	}
	294
	295	/*
	296	* We dropped and reaquired the lock, check
	297	* state again.
	298	*/
	299	if (!flist_empty(&sw->work_list))
	300	goto handle_work;
	301
	302	if (sw->flags & SW_F_EXIT) {
	303	pthread_mutex_unlock(&sw->lock);
	304	break;
	305	} else if (!(sw->flags & SW_F_IDLE)) {
	306	sw->flags \|= SW_F_IDLE;
	307	wq->next_free_worker = sw->index;
	308	if (wq->wake_idle)
	309	pthread_cond_signal(&wq->flush_cond);
	310	}
	311	update_accounting(sw);
	312	pthread_cond_wait(&sw->cond, &sw->lock);
	313	} else {
	314	handle_work:
	315	flist_splice_init(&sw->work_list, &local_list);
	316	}
	317	pthread_mutex_unlock(&sw->lock);
	318	handle_list(sw, &local_list);
	319	}
	320
	321	update_accounting(sw);
	322
	323	done:
	324	pthread_mutex_lock(&sw->lock);
	325	sw->flags \|= (SW_F_EXITED \| eflags);
	326	pthread_mutex_unlock(&sw->lock);
	327	return NULL;
	328	}
	329
	330	static void free_worker(struct submit_worker *sw)
	331	{
	332	struct thread_data *td = &sw->td;
	333
	334	fio_options_free(td);
	335	close_and_free_files(td);
	336	if (td->io_ops)
	337	close_ioengine(td);
	338	td_set_runstate(td, TD_EXITED);
	339
	340	pthread_cond_destroy(&sw->cond);
	341	pthread_mutex_destroy(&sw->lock);
	342	}
	343
	344	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	345	{
	346	struct thread_data *parent = sw->wq->td;
	347
	348	pthread_join(sw->thread, NULL);
	349	(*sum_cnt)++;
	350	sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt);
	351	free_worker(sw);
	352	}
	353
	354	void workqueue_exit(struct workqueue *wq)
	355	{
356	unsigned int shutdown, sum_cnt = 0;
357	struct submit_worker *sw;
358	int i;
359
360	for (i = 0; i < wq->max_workers; i++) {
361	sw = &wq->workers[i];
362
363	pthread_mutex_lock(&sw->lock);
364	sw->flags \|= SW_F_EXIT;
365	pthread_cond_signal(&sw->cond);
366	pthread_mutex_unlock(&sw->lock);
367	}
368
369	do {
370	shutdown = 0;
371	for (i = 0; i < wq->max_workers; i++) {
372	sw = &wq->workers[i];
373	if (sw->flags & SW_F_ACCOUNTED)
374	continue;
375	sw->flags \|= SW_F_ACCOUNTED;
376	shutdown_worker(sw, &sum_cnt);
377	shutdown++;
378	}
379	} while (shutdown && shutdown != wq->max_workers);
380
381	free(wq->workers);
382	pthread_mutex_destroy(&wq->flush_lock);
383	pthread_cond_destroy(&wq->flush_cond);
2a274336	384	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	385	}
	386
	387	static int start_worker(struct workqueue *wq, unsigned int index)
	388	{
	389	struct submit_worker *sw = &wq->workers[index];
	390	int ret;
	391
	392	INIT_FLIST_HEAD(&sw->work_list);
	393	pthread_cond_init(&sw->cond, NULL);
	394	pthread_mutex_init(&sw->lock, NULL);
	395	sw->wq = wq;
	396	sw->index = index;
	397
	398	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	399	if (!ret) {
	400	pthread_mutex_lock(&sw->lock);
	401	sw->flags = SW_F_IDLE;
	402	pthread_mutex_unlock(&sw->lock);
	403	return 0;
	404	}
	405
	406	free_worker(sw);
	407	return 1;
	408	}
	409
	410	int workqueue_init(struct thread_data td, struct workqueue wq,
	411	workqueue_fn *fn, unsigned max_pending)
	412	{
	413	unsigned int running;
	414	int i, error;
	415
	416	wq->max_workers = max_pending;
	417	wq->td = td;
	418	wq->fn = fn;
	419	wq->work_seq = 0;
	420	wq->next_free_worker = 0;
	421	pthread_cond_init(&wq->flush_cond, NULL);
	422	pthread_mutex_init(&wq->flush_lock, NULL);
2a274336	423	pthread_mutex_init(&wq->stat_lock, NULL);
a9da8ab2 JA	424
	425	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
	426
	427	for (i = 0; i < wq->max_workers; i++)
	428	if (start_worker(wq, i))
	429	break;
	430
	431	wq->max_workers = i;
bddc8d16 JA	432	if (!wq->max_workers)
bddc8d16 JA	433	goto err;
a9da8ab2 JA	434
	435	/*
	436	* Wait for them all to be started and initialized
	437	*/
	438	error = 0;
	439	do {
	440	struct submit_worker *sw;
	441
	442	running = 0;
	443	pthread_mutex_lock(&wq->flush_lock);
	444	for (i = 0; i < wq->max_workers; i++) {
	445	sw = &wq->workers[i];
	446	pthread_mutex_lock(&sw->lock);
	447	if (sw->flags & SW_F_RUNNING)
	448	running++;
	449	if (sw->flags & SW_F_ERROR)
	450	error++;
	451	pthread_mutex_unlock(&sw->lock);
	452	}
	453
	454	if (error \|\| running == wq->max_workers) {
	455	pthread_mutex_unlock(&wq->flush_lock);
	456	break;
	457	}
	458
	459	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	460	pthread_mutex_unlock(&wq->flush_lock);
	461	} while (1);
	462
bddc8d16 JA	463	if (!error)
bddc8d16 JA	464	return 0;
a9da8ab2	465
bddc8d16 JA	466	err:
	467	log_err("Can't create rate workqueue\n");
	468	td_verror(td, ESRCH, "workqueue_init");
	469	workqueue_exit(wq);
	470	return 1;
a9da8ab2	471	}