[fio.git] / workqueue.c

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

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

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 workqueue_work *work)
{
	struct submit_worker *sw;

	sw = get_submit_worker(wq);
	if (sw) {
		pthread_mutex_lock(&sw->lock);
		flist_add_tail(&work->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 workqueue_work *work;

	while (!flist_empty(list)) {
		work = flist_first_entry(list, struct workqueue_work, list);
		flist_del_init(&work->list);
		wq->ops.fn(sw, work);
	}
}

static void *worker_thread(void *data)
{
	struct submit_worker *sw = data;
	struct workqueue *wq = sw->wq;
	unsigned int eflags = 0, ret;
	FLIST_HEAD(local_list);

	ret = workqueue_init_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 (workqueue_pre_sleep_check(sw)) {
				pthread_mutex_unlock(&sw->lock);
				workqueue_pre_sleep(sw);
				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);
			}
			if (wq->ops.update_acct_fn)
				wq->ops.update_acct_fn(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);
	}

	if (wq->ops.update_acct_fn)
		wq->ops.update_acct_fn(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, unsigned int *sum_cnt)
{
	struct workqueue *wq = sw->wq;

	workqueue_exit_worker(sw, sum_cnt);

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

	if (wq->ops.free_worker_fn)
		wq->ops.free_worker_fn(sw);
}

static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
{
	pthread_join(sw->thread, NULL);
	free_worker(sw, sum_cnt);
}

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;

	if (wq->ops.alloc_worker_fn) {
		ret = wq->ops.alloc_worker_fn(sw);
		if (ret)
			return ret;
	}

	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, NULL);
	return 1;
}

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

	wq->max_workers = max_pending;
	wq->td = td;
	wq->ops = *ops;
	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"
a9da8ab2 JA	10	#include "flist.h"
	11	#include "workqueue.h"
	12	#include "lib/getrusage.h"
	13
a9da8ab2 JA	14	enum {
	15	SW_F_IDLE = 1 << 0,
	16	SW_F_RUNNING = 1 << 1,
	17	SW_F_EXIT = 1 << 2,
	18	SW_F_EXITED = 1 << 3,
	19	SW_F_ACCOUNTED = 1 << 4,
	20	SW_F_ERROR = 1 << 5,
	21	};
	22
	23	static struct submit_worker __get_submit_worker(struct workqueue wq,
	24	unsigned int start,
	25	unsigned int end,
	26	struct submit_worker **best)
	27	{
	28	struct submit_worker *sw = NULL;
	29
	30	while (start <= end) {
	31	sw = &wq->workers[start];
	32	if (sw->flags & SW_F_IDLE)
	33	return sw;
	34	if (!(best) \|\| sw->seq < (best)->seq)
	35	*best = sw;
	36	start++;
	37	}
	38
	39	return NULL;
	40	}
	41
	42	static struct submit_worker get_submit_worker(struct workqueue wq)
	43	{
	44	unsigned int next = wq->next_free_worker;
	45	struct submit_worker sw, best = NULL;
	46
	47	assert(next < wq->max_workers);
	48
	49	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	50	if (!sw && next)
	51	sw = __get_submit_worker(wq, 0, next - 1, &best);
	52
	53	/*
	54	* No truly idle found, use best match
	55	*/
	56	if (!sw)
	57	sw = best;
	58
	59	if (sw->index == wq->next_free_worker) {
	60	if (sw->index + 1 < wq->max_workers)
	61	wq->next_free_worker = sw->index + 1;
	62	else
	63	wq->next_free_worker = 0;
	64	}
	65
	66	return sw;
	67	}
	68
1391052a	69	static bool all_sw_idle(struct workqueue *wq)
a9da8ab2 JA	70	{
	71	int i;
	72
	73	for (i = 0; i < wq->max_workers; i++) {
	74	struct submit_worker *sw = &wq->workers[i];
	75
	76	if (!(sw->flags & SW_F_IDLE))
1391052a	77	return false;
a9da8ab2 JA	78	}
a9da8ab2 JA	79
1391052a	80	return true;
a9da8ab2 JA	81	}
	82
	83	/*
	84	* Must be serialized wrt workqueue_enqueue() by caller
	85	*/
	86	void workqueue_flush(struct workqueue *wq)
	87	{
	88	wq->wake_idle = 1;
	89
	90	while (!all_sw_idle(wq)) {
	91	pthread_mutex_lock(&wq->flush_lock);
	92	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	93	pthread_mutex_unlock(&wq->flush_lock);
	94	}
	95
	96	wq->wake_idle = 0;
	97	}
	98
	99	/*
b07f6ad1	100	* Must be serialized by caller. Returns true for queued, false for busy.
a9da8ab2	101	*/
88271841	102	bool workqueue_enqueue(struct workqueue wq, struct workqueue_work work)
a9da8ab2 JA	103	{
	104	struct submit_worker *sw;
	105
	106	sw = get_submit_worker(wq);
	107	if (sw) {
a9da8ab2	108	pthread_mutex_lock(&sw->lock);
88271841	109	flist_add_tail(&work->list, &sw->work_list);
a9da8ab2 JA	110	sw->seq = ++wq->work_seq;
	111	sw->flags &= ~SW_F_IDLE;
	112	pthread_mutex_unlock(&sw->lock);
	113
	114	pthread_cond_signal(&sw->cond);
b07f6ad1	115	return true;
a9da8ab2 JA	116	}
a9da8ab2 JA	117
b07f6ad1	118	return false;
a9da8ab2 JA	119	}
	120
	121	static void handle_list(struct submit_worker sw, struct flist_head list)
	122	{
	123	struct workqueue *wq = sw->wq;
88271841	124	struct workqueue_work *work;
a9da8ab2 JA	125
a9da8ab2 JA	126	while (!flist_empty(list)) {
88271841 JA	127	work = flist_first_entry(list, struct workqueue_work, list);
88271841 JA	128	flist_del_init(&work->list);
ee2b6d6e	129	wq->ops.fn(sw, work);
a9da8ab2 JA	130	}
	131	}
	132
a9da8ab2 JA	133	static void worker_thread(void data)
	134	{
	135	struct submit_worker *sw = data;
	136	struct workqueue *wq = sw->wq;
a9da8ab2 JA	137	unsigned int eflags = 0, ret;
	138	FLIST_HEAD(local_list);
	139
80ea26a8	140	ret = workqueue_init_worker(sw);
a9da8ab2 JA	141	pthread_mutex_lock(&sw->lock);
	142	sw->flags \|= SW_F_RUNNING;
	143	if (ret)
	144	sw->flags \|= SW_F_ERROR;
	145	pthread_mutex_unlock(&sw->lock);
	146
	147	pthread_mutex_lock(&wq->flush_lock);
	148	pthread_cond_signal(&wq->flush_cond);
	149	pthread_mutex_unlock(&wq->flush_lock);
	150
	151	if (sw->flags & SW_F_ERROR)
	152	goto done;
	153
	154	while (1) {
	155	pthread_mutex_lock(&sw->lock);
	156
	157	if (flist_empty(&sw->work_list)) {
	158	if (sw->flags & SW_F_EXIT) {
	159	pthread_mutex_unlock(&sw->lock);
	160	break;
	161	}
	162
ee2b6d6e	163	if (workqueue_pre_sleep_check(sw)) {
a9da8ab2	164	pthread_mutex_unlock(&sw->lock);
ee2b6d6e	165	workqueue_pre_sleep(sw);
a9da8ab2 JA	166	pthread_mutex_lock(&sw->lock);
	167	}
	168
	169	/*
	170	* We dropped and reaquired the lock, check
	171	* state again.
	172	*/
	173	if (!flist_empty(&sw->work_list))
	174	goto handle_work;
	175
	176	if (sw->flags & SW_F_EXIT) {
	177	pthread_mutex_unlock(&sw->lock);
	178	break;
	179	} else if (!(sw->flags & SW_F_IDLE)) {
	180	sw->flags \|= SW_F_IDLE;
	181	wq->next_free_worker = sw->index;
	182	if (wq->wake_idle)
	183	pthread_cond_signal(&wq->flush_cond);
	184	}
17ecadc3 JA	185	if (wq->ops.update_acct_fn)
	186	wq->ops.update_acct_fn(sw);
	187
a9da8ab2 JA	188	pthread_cond_wait(&sw->cond, &sw->lock);
	189	} else {
	190	handle_work:
	191	flist_splice_init(&sw->work_list, &local_list);
	192	}
	193	pthread_mutex_unlock(&sw->lock);
	194	handle_list(sw, &local_list);
	195	}
	196
17ecadc3 JA	197	if (wq->ops.update_acct_fn)
17ecadc3 JA	198	wq->ops.update_acct_fn(sw);
a9da8ab2 JA	199
	200	done:
	201	pthread_mutex_lock(&sw->lock);
	202	sw->flags \|= (SW_F_EXITED \| eflags);
	203	pthread_mutex_unlock(&sw->lock);
	204	return NULL;
	205	}
	206
c35c582d	207	static void free_worker(struct submit_worker sw, unsigned int sum_cnt)
a9da8ab2	208	{
ee2b6d6e	209	struct workqueue *wq = sw->wq;
a9da8ab2	210
c35c582d	211	workqueue_exit_worker(sw, sum_cnt);
a9da8ab2 JA	212
	213	pthread_cond_destroy(&sw->cond);
	214	pthread_mutex_destroy(&sw->lock);
ee2b6d6e JA	215
	216	if (wq->ops.free_worker_fn)
	217	wq->ops.free_worker_fn(sw);
a9da8ab2 JA	218	}
	219
	220	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	221	{
a9da8ab2	222	pthread_join(sw->thread, NULL);
c35c582d	223	free_worker(sw, sum_cnt);
a9da8ab2 JA	224	}
	225
	226	void workqueue_exit(struct workqueue *wq)
	227	{
	228	unsigned int shutdown, sum_cnt = 0;
	229	struct submit_worker *sw;
	230	int i;
	231
	232	for (i = 0; i < wq->max_workers; i++) {
	233	sw = &wq->workers[i];
	234
	235	pthread_mutex_lock(&sw->lock);
	236	sw->flags \|= SW_F_EXIT;
	237	pthread_cond_signal(&sw->cond);
	238	pthread_mutex_unlock(&sw->lock);
	239	}
	240
	241	do {
	242	shutdown = 0;
	243	for (i = 0; i < wq->max_workers; i++) {
	244	sw = &wq->workers[i];
	245	if (sw->flags & SW_F_ACCOUNTED)
	246	continue;
b7d0bbf1	247	pthread_mutex_lock(&sw->lock);
a9da8ab2	248	sw->flags \|= SW_F_ACCOUNTED;
b7d0bbf1	249	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	250	shutdown_worker(sw, &sum_cnt);
	251	shutdown++;
	252	}
	253	} while (shutdown && shutdown != wq->max_workers);
	254
	255	free(wq->workers);
	256	pthread_mutex_destroy(&wq->flush_lock);
	257	pthread_cond_destroy(&wq->flush_cond);
2a274336	258	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	259	}
	260
	261	static int start_worker(struct workqueue *wq, unsigned int index)
	262	{
	263	struct submit_worker *sw = &wq->workers[index];
	264	int ret;
	265
	266	INIT_FLIST_HEAD(&sw->work_list);
	267	pthread_cond_init(&sw->cond, NULL);
	268	pthread_mutex_init(&sw->lock, NULL);
	269	sw->wq = wq;
	270	sw->index = index;
	271
ee2b6d6e JA	272	if (wq->ops.alloc_worker_fn) {
	273	ret = wq->ops.alloc_worker_fn(sw);
	274	if (ret)
	275	return ret;
	276	}
	277
a9da8ab2 JA	278	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	279	if (!ret) {
	280	pthread_mutex_lock(&sw->lock);
	281	sw->flags = SW_F_IDLE;
	282	pthread_mutex_unlock(&sw->lock);
	283	return 0;
	284	}
	285
c35c582d	286	free_worker(sw, NULL);
a9da8ab2 JA	287	return 1;
	288	}
	289
	290	int workqueue_init(struct thread_data td, struct workqueue wq,
5bb79f69	291	struct workqueue_ops *ops, unsigned max_pending)
a9da8ab2 JA	292	{
	293	unsigned int running;
	294	int i, error;
	295
	296	wq->max_workers = max_pending;
	297	wq->td = td;
5bb79f69	298	wq->ops = *ops;
a9da8ab2 JA	299	wq->work_seq = 0;
	300	wq->next_free_worker = 0;
	301	pthread_cond_init(&wq->flush_cond, NULL);
	302	pthread_mutex_init(&wq->flush_lock, NULL);
2a274336	303	pthread_mutex_init(&wq->stat_lock, NULL);
a9da8ab2 JA	304
	305	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
	306
	307	for (i = 0; i < wq->max_workers; i++)
	308	if (start_worker(wq, i))
	309	break;
	310
	311	wq->max_workers = i;
bddc8d16 JA	312	if (!wq->max_workers)
bddc8d16 JA	313	goto err;
a9da8ab2 JA	314
	315	/*
	316	* Wait for them all to be started and initialized
	317	*/
	318	error = 0;
	319	do {
	320	struct submit_worker *sw;
	321
	322	running = 0;
	323	pthread_mutex_lock(&wq->flush_lock);
	324	for (i = 0; i < wq->max_workers; i++) {
	325	sw = &wq->workers[i];
	326	pthread_mutex_lock(&sw->lock);
	327	if (sw->flags & SW_F_RUNNING)
	328	running++;
	329	if (sw->flags & SW_F_ERROR)
	330	error++;
	331	pthread_mutex_unlock(&sw->lock);
	332	}
	333
	334	if (error \|\| running == wq->max_workers) {
	335	pthread_mutex_unlock(&wq->flush_lock);
	336	break;
	337	}
	338
	339	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	340	pthread_mutex_unlock(&wq->flush_lock);
	341	} while (1);
	342
bddc8d16 JA	343	if (!error)
bddc8d16 JA	344	return 0;
a9da8ab2	345
bddc8d16 JA	346	err:
	347	log_err("Can't create rate workqueue\n");
	348	td_verror(td, ESRCH, "workqueue_init");
	349	workqueue_exit(wq);
	350	return 1;
a9da8ab2	351	}