[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_ACCOUNTED	= 1 << 3,
	SW_F_ERROR	= 1 << 4,
};

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 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:
	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);

	ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
	if (ret)
		return ret;

	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;
	int ret;

	wq->max_workers = max_workers;
	wq->td = td;
	wq->ops = *ops;
	wq->work_seq = 0;
	wq->next_free_worker = 0;

	ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
	if (ret)
		goto err;
	ret = mutex_init_pshared(&wq->stat_lock);
	if (ret)
		goto err;

	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
	if (!wq->workers)
		goto err;

	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,
d22042d2 JA	18	SW_F_ACCOUNTED = 1 << 3,
d22042d2 JA	19	SW_F_ERROR = 1 << 4,
a9da8ab2 JA	20	};
	21
	22	static struct submit_worker __get_submit_worker(struct workqueue wq,
	23	unsigned int start,
	24	unsigned int end,
	25	struct submit_worker **best)
	26	{
	27	struct submit_worker *sw = NULL;
	28
	29	while (start <= end) {
	30	sw = &wq->workers[start];
	31	if (sw->flags & SW_F_IDLE)
	32	return sw;
	33	if (!(best) \|\| sw->seq < (best)->seq)
	34	*best = sw;
	35	start++;
	36	}
	37
	38	return NULL;
	39	}
	40
	41	static struct submit_worker get_submit_worker(struct workqueue wq)
	42	{
	43	unsigned int next = wq->next_free_worker;
	44	struct submit_worker sw, best = NULL;
	45
	46	assert(next < wq->max_workers);
	47
	48	sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best);
	49	if (!sw && next)
	50	sw = __get_submit_worker(wq, 0, next - 1, &best);
	51
	52	/*
	53	* No truly idle found, use best match
	54	*/
	55	if (!sw)
	56	sw = best;
	57
	58	if (sw->index == wq->next_free_worker) {
	59	if (sw->index + 1 < wq->max_workers)
	60	wq->next_free_worker = sw->index + 1;
	61	else
	62	wq->next_free_worker = 0;
	63	}
	64
	65	return sw;
	66	}
	67
1391052a	68	static bool all_sw_idle(struct workqueue *wq)
a9da8ab2 JA	69	{
	70	int i;
	71
	72	for (i = 0; i < wq->max_workers; i++) {
	73	struct submit_worker *sw = &wq->workers[i];
	74
	75	if (!(sw->flags & SW_F_IDLE))
1391052a	76	return false;
a9da8ab2 JA	77	}
a9da8ab2 JA	78
1391052a	79	return true;
a9da8ab2 JA	80	}
	81
	82	/*
	83	* Must be serialized wrt workqueue_enqueue() by caller
	84	*/
	85	void workqueue_flush(struct workqueue *wq)
	86	{
	87	wq->wake_idle = 1;
	88
	89	while (!all_sw_idle(wq)) {
	90	pthread_mutex_lock(&wq->flush_lock);
	91	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	92	pthread_mutex_unlock(&wq->flush_lock);
	93	}
	94
	95	wq->wake_idle = 0;
	96	}
	97
	98	/*
b07f6ad1	99	* Must be serialized by caller. Returns true for queued, false for busy.
a9da8ab2	100	*/
26de50cf	101	void workqueue_enqueue(struct workqueue wq, struct workqueue_work work)
a9da8ab2 JA	102	{
	103	struct submit_worker *sw;
	104
	105	sw = get_submit_worker(wq);
26de50cf JA	106	assert(sw);
	107
	108	pthread_mutex_lock(&sw->lock);
	109	flist_add_tail(&work->list, &sw->work_list);
	110	sw->seq = ++wq->work_seq;
	111	sw->flags &= ~SW_F_IDLE;
	112	pthread_mutex_unlock(&sw->lock);
a9da8ab2	113
26de50cf	114	pthread_cond_signal(&sw->cond);
a9da8ab2 JA	115	}
	116
	117	static void handle_list(struct submit_worker sw, struct flist_head list)
	118	{
	119	struct workqueue *wq = sw->wq;
88271841	120	struct workqueue_work *work;
a9da8ab2 JA	121
a9da8ab2 JA	122	while (!flist_empty(list)) {
88271841 JA	123	work = flist_first_entry(list, struct workqueue_work, list);
88271841 JA	124	flist_del_init(&work->list);
ee2b6d6e	125	wq->ops.fn(sw, work);
a9da8ab2 JA	126	}
	127	}
	128
a9da8ab2 JA	129	static void worker_thread(void data)
	130	{
	131	struct submit_worker *sw = data;
	132	struct workqueue *wq = sw->wq;
d22042d2	133	unsigned int ret = 0;
a9da8ab2 JA	134	FLIST_HEAD(local_list);
a9da8ab2 JA	135
24660963 JA	136	sk_out_assign(sw->sk_out);
24660963 JA	137
f6496ba7 JA	138	if (wq->ops.nice) {
	139	if (nice(wq->ops.nice) < 0) {
	140	log_err("workqueue: nice %s\n", strerror(errno));
	141	ret = 1;
	142	}
	143	}
	144
	145	if (!ret)
	146	ret = workqueue_init_worker(sw);
	147
a9da8ab2 JA	148	pthread_mutex_lock(&sw->lock);
	149	sw->flags \|= SW_F_RUNNING;
	150	if (ret)
	151	sw->flags \|= SW_F_ERROR;
	152	pthread_mutex_unlock(&sw->lock);
	153
	154	pthread_mutex_lock(&wq->flush_lock);
	155	pthread_cond_signal(&wq->flush_cond);
	156	pthread_mutex_unlock(&wq->flush_lock);
	157
	158	if (sw->flags & SW_F_ERROR)
	159	goto done;
	160
	161	while (1) {
	162	pthread_mutex_lock(&sw->lock);
	163
	164	if (flist_empty(&sw->work_list)) {
	165	if (sw->flags & SW_F_EXIT) {
	166	pthread_mutex_unlock(&sw->lock);
	167	break;
	168	}
	169
ee2b6d6e	170	if (workqueue_pre_sleep_check(sw)) {
a9da8ab2	171	pthread_mutex_unlock(&sw->lock);
ee2b6d6e	172	workqueue_pre_sleep(sw);
a9da8ab2 JA	173	pthread_mutex_lock(&sw->lock);
	174	}
	175
	176	/*
	177	* We dropped and reaquired the lock, check
	178	* state again.
	179	*/
	180	if (!flist_empty(&sw->work_list))
	181	goto handle_work;
	182
	183	if (sw->flags & SW_F_EXIT) {
	184	pthread_mutex_unlock(&sw->lock);
	185	break;
	186	} else if (!(sw->flags & SW_F_IDLE)) {
	187	sw->flags \|= SW_F_IDLE;
	188	wq->next_free_worker = sw->index;
	189	if (wq->wake_idle)
	190	pthread_cond_signal(&wq->flush_cond);
	191	}
17ecadc3 JA	192	if (wq->ops.update_acct_fn)
	193	wq->ops.update_acct_fn(sw);
	194
a9da8ab2 JA	195	pthread_cond_wait(&sw->cond, &sw->lock);
	196	} else {
	197	handle_work:
	198	flist_splice_init(&sw->work_list, &local_list);
	199	}
	200	pthread_mutex_unlock(&sw->lock);
	201	handle_list(sw, &local_list);
	202	}
	203
17ecadc3 JA	204	if (wq->ops.update_acct_fn)
17ecadc3 JA	205	wq->ops.update_acct_fn(sw);
a9da8ab2 JA	206
a9da8ab2 JA	207	done:
24660963	208	sk_out_drop();
a9da8ab2 JA	209	return NULL;
	210	}
	211
c35c582d	212	static void free_worker(struct submit_worker sw, unsigned int sum_cnt)
a9da8ab2	213	{
ee2b6d6e	214	struct workqueue *wq = sw->wq;
a9da8ab2	215
c35c582d	216	workqueue_exit_worker(sw, sum_cnt);
a9da8ab2 JA	217
	218	pthread_cond_destroy(&sw->cond);
	219	pthread_mutex_destroy(&sw->lock);
ee2b6d6e JA	220
	221	if (wq->ops.free_worker_fn)
	222	wq->ops.free_worker_fn(sw);
a9da8ab2 JA	223	}
	224
	225	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
	226	{
a9da8ab2	227	pthread_join(sw->thread, NULL);
c35c582d	228	free_worker(sw, sum_cnt);
a9da8ab2 JA	229	}
	230
	231	void workqueue_exit(struct workqueue *wq)
	232	{
	233	unsigned int shutdown, sum_cnt = 0;
	234	struct submit_worker *sw;
	235	int i;
	236
f8e266c9 JA	237	if (!wq->workers)
	238	return;
	239
a9da8ab2 JA	240	for (i = 0; i < wq->max_workers; i++) {
	241	sw = &wq->workers[i];
	242
	243	pthread_mutex_lock(&sw->lock);
	244	sw->flags \|= SW_F_EXIT;
	245	pthread_cond_signal(&sw->cond);
	246	pthread_mutex_unlock(&sw->lock);
	247	}
	248
	249	do {
	250	shutdown = 0;
	251	for (i = 0; i < wq->max_workers; i++) {
	252	sw = &wq->workers[i];
	253	if (sw->flags & SW_F_ACCOUNTED)
	254	continue;
b7d0bbf1	255	pthread_mutex_lock(&sw->lock);
a9da8ab2	256	sw->flags \|= SW_F_ACCOUNTED;
b7d0bbf1	257	pthread_mutex_unlock(&sw->lock);
a9da8ab2 JA	258	shutdown_worker(sw, &sum_cnt);
	259	shutdown++;
	260	}
	261	} while (shutdown && shutdown != wq->max_workers);
	262
a47591e4	263	sfree(wq->workers);
f8e266c9	264	wq->workers = NULL;
a9da8ab2 JA	265	pthread_mutex_destroy(&wq->flush_lock);
a9da8ab2 JA	266	pthread_cond_destroy(&wq->flush_cond);
2a274336	267	pthread_mutex_destroy(&wq->stat_lock);
a9da8ab2 JA	268	}
a9da8ab2 JA	269
24660963 JA	270	static int start_worker(struct workqueue *wq, unsigned int index,
24660963 JA	271	struct sk_out *sk_out)
a9da8ab2 JA	272	{
	273	struct submit_worker *sw = &wq->workers[index];
	274	int ret;
	275
	276	INIT_FLIST_HEAD(&sw->work_list);
34febb23 JA	277
34febb23 JA	278	ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
f9e5b5ee JK	279	if (ret)
f9e5b5ee JK	280	return ret;
34febb23	281
a9da8ab2 JA	282	sw->wq = wq;
a9da8ab2 JA	283	sw->index = index;
24660963	284	sw->sk_out = sk_out;
a9da8ab2	285
ee2b6d6e JA	286	if (wq->ops.alloc_worker_fn) {
	287	ret = wq->ops.alloc_worker_fn(sw);
	288	if (ret)
	289	return ret;
	290	}
	291
a9da8ab2 JA	292	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
	293	if (!ret) {
	294	pthread_mutex_lock(&sw->lock);
	295	sw->flags = SW_F_IDLE;
	296	pthread_mutex_unlock(&sw->lock);
	297	return 0;
	298	}
	299
c35c582d	300	free_worker(sw, NULL);
a9da8ab2 JA	301	return 1;
	302	}
	303
	304	int workqueue_init(struct thread_data td, struct workqueue wq,
24660963 JA	305	struct workqueue_ops *ops, unsigned int max_workers,
24660963 JA	306	struct sk_out *sk_out)
a9da8ab2 JA	307	{
	308	unsigned int running;
	309	int i, error;
f9e5b5ee	310	int ret;
a9da8ab2	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;
a9da8ab2 JA	316	wq->next_free_worker = 0;
f9e5b5ee	317
34febb23 JA	318	ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
34febb23 JA	319	if (ret)
f9e5b5ee	320	goto err;
34febb23 JA	321	ret = mutex_init_pshared(&wq->stat_lock);
34febb23 JA	322	if (ret)
f9e5b5ee	323	goto err;
a9da8ab2	324
a47591e4	325	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
81b3c86f JA	326	if (!wq->workers)
81b3c86f JA	327	goto err;
a9da8ab2 JA	328
a9da8ab2 JA	329	for (i = 0; i < wq->max_workers; i++)
24660963	330	if (start_worker(wq, i, sk_out))
a9da8ab2 JA	331	break;
	332
	333	wq->max_workers = i;
bddc8d16 JA	334	if (!wq->max_workers)
bddc8d16 JA	335	goto err;
a9da8ab2 JA	336
	337	/*
	338	* Wait for them all to be started and initialized
	339	*/
	340	error = 0;
	341	do {
	342	struct submit_worker *sw;
	343
	344	running = 0;
	345	pthread_mutex_lock(&wq->flush_lock);
	346	for (i = 0; i < wq->max_workers; i++) {
	347	sw = &wq->workers[i];
	348	pthread_mutex_lock(&sw->lock);
	349	if (sw->flags & SW_F_RUNNING)
	350	running++;
	351	if (sw->flags & SW_F_ERROR)
	352	error++;
	353	pthread_mutex_unlock(&sw->lock);
	354	}
	355
	356	if (error \|\| running == wq->max_workers) {
	357	pthread_mutex_unlock(&wq->flush_lock);
	358	break;
	359	}
	360
	361	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
	362	pthread_mutex_unlock(&wq->flush_lock);
	363	} while (1);
	364
bddc8d16 JA	365	if (!error)
bddc8d16 JA	366	return 0;
a9da8ab2	367
bddc8d16 JA	368	err:
	369	log_err("Can't create rate workqueue\n");
	370	td_verror(td, ESRCH, "workqueue_init");
	371	workqueue_exit(wq);
	372	return 1;
a9da8ab2	373	}