[fio.git] / workqueue.c

/*
 * Generic workqueue offload mechanism
 *
 * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
 *
 */
#include <unistd.h>

#include "fio.h"
#include "flist.h"
#include "workqueue.h"
#include "smalloc.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.
 */
void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
{
	struct submit_worker *sw;

	sw = get_submit_worker(wq);
	assert(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);
}

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 = 0;
	FLIST_HEAD(local_list);

	sk_out_assign(sw->sk_out);

	if (wq->ops.nice) {
		if (nice(wq->ops.nice) < 0) {
			log_err("workqueue: nice %s\n", strerror(errno));
			ret = 1;
		}
	}

	if (!ret)
		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);
	sk_out_drop();
	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;

	if (!wq->workers)
		return;

	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);

	sfree(wq->workers);
	wq->workers = NULL;
	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 sk_out *sk_out)
{
	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;
	sw->sk_out = sk_out;

	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 int max_workers,
		   struct sk_out *sk_out)
{
	unsigned int running;
	int i, error;

	wq->max_workers = max_workers;
	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 = smalloc(wq->max_workers * sizeof(struct submit_worker));

	for (i = 0; i < wq->max_workers; i++)
		if (start_worker(wq, i, sk_out))
			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	1	/*
51575029	2	* Generic workqueue offload mechanism
a9da8ab2 JA	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"
a9da8ab2 JA	11	#include "workqueue.h"
a47591e4	12	#include "smalloc.h"
a9da8ab2	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	*/
26de50cf	102	void workqueue_enqueue(struct workqueue wq, struct workqueue_work work)
a9da8ab2 JA	103	{
	104	struct submit_worker *sw;
	105
	106	sw = get_submit_worker(wq);
26de50cf JA	107	assert(sw);
	108
	109	pthread_mutex_lock(&sw->lock);
	110	flist_add_tail(&work->list, &sw->work_list);
	111	sw->seq = ++wq->work_seq;
	112	sw->flags &= ~SW_F_IDLE;
	113	pthread_mutex_unlock(&sw->lock);
a9da8ab2	114
26de50cf	115	pthread_cond_signal(&sw->cond);
a9da8ab2 JA	116	}
	117
	118	static void handle_list(struct submit_worker sw, struct flist_head list)
	119	{
	120	struct workqueue *wq = sw->wq;
88271841	121	struct workqueue_work *work;
a9da8ab2 JA	122
a9da8ab2 JA	123	while (!flist_empty(list)) {
88271841 JA	124	work = flist_first_entry(list, struct workqueue_work, list);
88271841 JA	125	flist_del_init(&work->list);
ee2b6d6e	126	wq->ops.fn(sw, work);
a9da8ab2 JA	127	}
	128	}
	129
a9da8ab2 JA	130	static void worker_thread(void data)
	131	{
	132	struct submit_worker *sw = data;
	133	struct workqueue *wq = sw->wq;
f6496ba7	134	unsigned int eflags = 0, ret = 0;
a9da8ab2 JA	135	FLIST_HEAD(local_list);
a9da8ab2 JA	136
24660963 JA	137	sk_out_assign(sw->sk_out);
24660963 JA	138
f6496ba7 JA	139	if (wq->ops.nice) {
	140	if (nice(wq->ops.nice) < 0) {
	141	log_err("workqueue: nice %s\n", strerror(errno));
	142	ret = 1;
	143	}
	144	}
	145
	146	if (!ret)
	147	ret = workqueue_init_worker(sw);
	148
a9da8ab2 JA	149	pthread_mutex_lock(&sw->lock);
	150	sw->flags \|= SW_F_RUNNING;
	151	if (ret)
	152	sw->flags \|= SW_F_ERROR;
	153	pthread_mutex_unlock(&sw->lock);
	154
	155	pthread_mutex_lock(&wq->flush_lock);
	156	pthread_cond_signal(&wq->flush_cond);
	157	pthread_mutex_unlock(&wq->flush_lock);
	158
	159	if (sw->flags & SW_F_ERROR)
	160	goto done;
	161
	162	while (1) {
	163	pthread_mutex_lock(&sw->lock);
	164
	165	if (flist_empty(&sw->work_list)) {
	166	if (sw->flags & SW_F_EXIT) {
	167	pthread_mutex_unlock(&sw->lock);
	168	break;
	169	}
	170
ee2b6d6e	171	if (workqueue_pre_sleep_check(sw)) {
a9da8ab2	172	pthread_mutex_unlock(&sw->lock);
ee2b6d6e	173	workqueue_pre_sleep(sw);
a9da8ab2 JA	174	pthread_mutex_lock(&sw->lock);
	175	}
	176
	177	/*
	178	* We dropped and reaquired the lock, check
	179	* state again.
	180	*/
	181	if (!flist_empty(&sw->work_list))
	182	goto handle_work;
	183
	184	if (sw->flags & SW_F_EXIT) {
	185	pthread_mutex_unlock(&sw->lock);
	186	break;
	187	} else if (!(sw->flags & SW_F_IDLE)) {
	188	sw->flags \|= SW_F_IDLE;
	189	wq->next_free_worker = sw->index;
	190	if (wq->wake_idle)
	191	pthread_cond_signal(&wq->flush_cond);
	192	}
17ecadc3 JA	193	if (wq->ops.update_acct_fn)
	194	wq->ops.update_acct_fn(sw);
	195
a9da8ab2 JA	196	pthread_cond_wait(&sw->cond, &sw->lock);
	197	} else {
	198	handle_work:
	199	flist_splice_init(&sw->work_list, &local_list);
	200	}
	201	pthread_mutex_unlock(&sw->lock);
	202	handle_list(sw, &local_list);
	203	}
	204
17ecadc3 JA	205	if (wq->ops.update_acct_fn)
17ecadc3 JA	206	wq->ops.update_acct_fn(sw);
a9da8ab2 JA	207
	208	done:
	209	pthread_mutex_lock(&sw->lock);
	210	sw->flags \|= (SW_F_EXITED \| eflags);
	211	pthread_mutex_unlock(&sw->lock);
24660963	212	sk_out_drop();
a9da8ab2 JA	213	return NULL;
	214	}
	215
c35c582d	216	static void free_worker(struct submit_worker sw, unsigned int sum_cnt)
a9da8ab2	217	{
ee2b6d6e	218	struct workqueue *wq = sw->wq;
a9da8ab2	219
c35c582d	220	workqueue_exit_worker(sw, sum_cnt);
a9da8ab2 JA	221
	222	pthread_cond_destroy(&sw->cond);
	223	pthread_mutex_destroy(&sw->lock);
ee2b6d6e JA	224
	225	if (wq->ops.free_worker_fn)
	226	wq->ops.free_worker_fn(sw);
a9da8ab2 JA	227	}
	228
	229	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	230	{
a9da8ab2	231	pthread_join(sw->thread, NULL);
c35c582d	232	free_worker(sw, sum_cnt);
a9da8ab2 JA	233	}
	234
	235	void workqueue_exit(struct workqueue *wq)
	236	{
	237	unsigned int shutdown, sum_cnt = 0;
	238	struct submit_worker *sw;
	239	int i;
	240
f8e266c9 JA	241	if (!wq->workers)
	242	return;
	243
a9da8ab2 JA	244	for (i = 0; i < wq->max_workers; i++) {
	245	sw = &wq->workers[i];
	246
	247	pthread_mutex_lock(&sw->lock);
	248	sw->flags \|= SW_F_EXIT;
	249	pthread_cond_signal(&sw->cond);
	250	pthread_mutex_unlock(&sw->lock);
	251	}
	252
	253	do {
	254	shutdown = 0;
	255	for (i = 0; i < wq->max_workers; i++) {
	256	sw = &wq->workers[i];
	257	if (sw->flags & SW_F_ACCOUNTED)
	258	continue;
b7d0bbf1	259	pthread_mutex_lock(&sw->lock);
a9da8ab2	260	sw->flags \|= SW_F_ACCOUNTED;
b7d0bbf1	261	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	262	shutdown_worker(sw, &sum_cnt);
	263	shutdown++;
	264	}
	265	} while (shutdown && shutdown != wq->max_workers);
	266
a47591e4	267	sfree(wq->workers);
f8e266c9	268	wq->workers = NULL;
a9da8ab2 JA	269	pthread_mutex_destroy(&wq->flush_lock);
a9da8ab2 JA	270	pthread_cond_destroy(&wq->flush_cond);
2a274336	271	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	272	}
a9da8ab2 JA	273
24660963 JA	274	static int start_worker(struct workqueue *wq, unsigned int index,
24660963 JA	275	struct sk_out *sk_out)
a9da8ab2 JA	276	{
	277	struct submit_worker *sw = &wq->workers[index];
	278	int ret;
	279
	280	INIT_FLIST_HEAD(&sw->work_list);
	281	pthread_cond_init(&sw->cond, NULL);
	282	pthread_mutex_init(&sw->lock, NULL);
	283	sw->wq = wq;
	284	sw->index = index;
24660963	285	sw->sk_out = sk_out;
a9da8ab2	286
ee2b6d6e JA	287	if (wq->ops.alloc_worker_fn) {
	288	ret = wq->ops.alloc_worker_fn(sw);
	289	if (ret)
	290	return ret;
	291	}
	292
a9da8ab2 JA	293	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	294	if (!ret) {
	295	pthread_mutex_lock(&sw->lock);
	296	sw->flags = SW_F_IDLE;
	297	pthread_mutex_unlock(&sw->lock);
	298	return 0;
	299	}
	300
c35c582d	301	free_worker(sw, NULL);
a9da8ab2 JA	302	return 1;
	303	}
	304
	305	int workqueue_init(struct thread_data td, struct workqueue wq,
24660963 JA	306	struct workqueue_ops *ops, unsigned int max_workers,
24660963 JA	307	struct sk_out *sk_out)
a9da8ab2 JA	308	{
	309	unsigned int running;
	310	int i, error;
	311
24660963	312	wq->max_workers = max_workers;
a9da8ab2	313	wq->td = td;
5bb79f69	314	wq->ops = *ops;
a9da8ab2 JA	315	wq->work_seq = 0;
	316	wq->next_free_worker = 0;
	317	pthread_cond_init(&wq->flush_cond, NULL);
	318	pthread_mutex_init(&wq->flush_lock, NULL);
2a274336	319	pthread_mutex_init(&wq->stat_lock, NULL);
a9da8ab2	320
a47591e4	321	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
a9da8ab2 JA	322
a9da8ab2 JA	323	for (i = 0; i < wq->max_workers; i++)
24660963	324	if (start_worker(wq, i, sk_out))
a9da8ab2 JA	325	break;
	326
	327	wq->max_workers = i;
bddc8d16 JA	328	if (!wq->max_workers)
bddc8d16 JA	329	goto err;
a9da8ab2 JA	330
	331	/*
	332	* Wait for them all to be started and initialized
	333	*/
	334	error = 0;
	335	do {
	336	struct submit_worker *sw;
	337
	338	running = 0;
	339	pthread_mutex_lock(&wq->flush_lock);
	340	for (i = 0; i < wq->max_workers; i++) {
	341	sw = &wq->workers[i];
	342	pthread_mutex_lock(&sw->lock);
	343	if (sw->flags & SW_F_RUNNING)
	344	running++;
	345	if (sw->flags & SW_F_ERROR)
	346	error++;
	347	pthread_mutex_unlock(&sw->lock);
	348	}
	349
	350	if (error \|\| running == wq->max_workers) {
	351	pthread_mutex_unlock(&wq->flush_lock);
	352	break;
	353	}
	354
	355	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	356	pthread_mutex_unlock(&wq->flush_lock);
	357	} while (1);
	358
bddc8d16 JA	359	if (!error)
bddc8d16 JA	360	return 0;
a9da8ab2	361
bddc8d16 JA	362	err:
	363	log_err("Can't create rate workqueue\n");
	364	td_verror(td, ESRCH, "workqueue_init");
	365	workqueue_exit(wq);
	366	return 1;
a9da8ab2	367	}