[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 int 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 0;
	}

	return 1;
}

/*
 * 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;
		}

		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);
	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);

	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;
}

static void sum_val(uint64_t *dst, uint64_t *src)
{
	if (*src) {
		__sync_fetch_and_add(dst, *src);
		*src = 0;
	}
}

static void sum_ddir(struct thread_data *dst, struct thread_data *src,
		     enum fio_ddir ddir)
{
	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]);
}

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) {
				pthread_mutex_unlock(&sw->lock);
				io_u_quiesce(td);
				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);
}

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);

	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) {
err:
		log_err("Can't create rate workqueue\n");
		td_verror(td, ESRCH, "workqueue_init");
		workqueue_exit(wq);
		return 1;
	}

	/*
	 * 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)
		goto err;

	return 0;
}
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
82	static int all_sw_idle(struct workqueue *wq)
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))
90	return 0;
91	}
92
93	return 1;
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;
127	}
128
129	pthread_mutex_lock(&sw->lock);
130	flist_add_tail(&io_u->verify_list, &sw->work_list);
131	sw->seq = ++wq->work_seq;
132	sw->flags &= ~SW_F_IDLE;
133	pthread_mutex_unlock(&sw->lock);
134
135	pthread_cond_signal(&sw->cond);
136	return FIO_Q_QUEUED;
137	}
138
139	return FIO_Q_BUSY;
140	}
141
142	static void handle_list(struct submit_worker sw, struct flist_head list)
143	{
144	struct workqueue *wq = sw->wq;
145	struct io_u *io_u;
146
147	while (!flist_empty(list)) {
148	io_u = flist_first_entry(list, struct io_u, verify_list);
149	flist_del_init(&io_u->verify_list);
150	wq->fn(&sw->td, io_u);
151	}
152	}
153
154	static int init_submit_worker(struct submit_worker *sw)
155	{
156	struct thread_data *parent = sw->wq->td;
157	struct thread_data *td = &sw->td;
158	int fio_unused ret;
159
160	memcpy(&td->o, &parent->o, sizeof(td->o));
161	memcpy(&td->ts, &parent->ts, sizeof(td->ts));
162	td->o.uid = td->o.gid = -1U;
163	dup_files(td, parent);
164	fio_options_mem_dupe(td);
165
166	if (ioengine_load(td))
167	goto err;
168
169	if (td->o.odirect)
170	td->io_ops->flags \|= FIO_RAWIO;
171
172	td->pid = gettid();
173
174	INIT_FLIST_HEAD(&td->io_log_list);
175	INIT_FLIST_HEAD(&td->io_hist_list);
176	INIT_FLIST_HEAD(&td->verify_list);
177	INIT_FLIST_HEAD(&td->trim_list);
178	INIT_FLIST_HEAD(&td->next_rand_list);
179	td->io_hist_tree = RB_ROOT;
180
181	td->o.iodepth = 1;
182	if (td_io_init(td))
183	goto err_io_init;
184
185	fio_gettime(&td->epoch, NULL);
186	fio_getrusage(&td->ru_start);
187	clear_io_state(td);
188
189	td_set_runstate(td, TD_RUNNING);
190	td->flags \|= TD_F_CHILD;
191	td->parent = parent;
192	return 0;
193
194	err_io_init:
195	close_ioengine(td);
196	err:
197	return 1;
198	}
199
200	static void sum_val(uint64_t dst, uint64_t src)
201	{
202	if (*src) {
203	__sync_fetch_and_add(dst, *src);
204	*src = 0;
205	}
206	}
207
208	static void sum_ddir(struct thread_data dst, struct thread_data src,
209	enum fio_ddir ddir)
210	{
211	sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
212	sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
213	sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
214	sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
215	sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
216	}
217
218	static void update_accounting(struct submit_worker *sw)
219	{
220	struct thread_data *src = &sw->td;
221	struct thread_data *dst = sw->wq->td;
222
223	if (td_read(src))
224	sum_ddir(dst, src, DDIR_READ);
225	if (td_write(src))
226	sum_ddir(dst, src, DDIR_WRITE);
227	if (td_trim(src))
228	sum_ddir(dst, src, DDIR_TRIM);
229	}
230
231	static void worker_thread(void data)
232	{
233	struct submit_worker *sw = data;
234	struct workqueue *wq = sw->wq;
235	struct thread_data *td = &sw->td;
236	unsigned int eflags = 0, ret;
237	FLIST_HEAD(local_list);
238
239	ret = init_submit_worker(sw);
240	pthread_mutex_lock(&sw->lock);
241	sw->flags \|= SW_F_RUNNING;
242	if (ret)
243	sw->flags \|= SW_F_ERROR;
244	pthread_mutex_unlock(&sw->lock);
245
246	pthread_mutex_lock(&wq->flush_lock);
247	pthread_cond_signal(&wq->flush_cond);
248	pthread_mutex_unlock(&wq->flush_lock);
249
250	if (sw->flags & SW_F_ERROR)
251	goto done;
252
253	while (1) {
254	pthread_mutex_lock(&sw->lock);
255
256	if (flist_empty(&sw->work_list)) {
257	if (sw->flags & SW_F_EXIT) {
258	pthread_mutex_unlock(&sw->lock);
259	break;
260	}
261
262	if (td->io_u_queued \|\| td->cur_depth \|\|
263	td->io_u_in_flight) {
264	pthread_mutex_unlock(&sw->lock);
265	io_u_quiesce(td);
266	pthread_mutex_lock(&sw->lock);
267	}
268
269	/*
270	* We dropped and reaquired the lock, check
271	* state again.
272	*/
273	if (!flist_empty(&sw->work_list))
274	goto handle_work;
275
276	if (sw->flags & SW_F_EXIT) {
277	pthread_mutex_unlock(&sw->lock);
278	break;
279	} else if (!(sw->flags & SW_F_IDLE)) {
280	sw->flags \|= SW_F_IDLE;
281	wq->next_free_worker = sw->index;
282	if (wq->wake_idle)
283	pthread_cond_signal(&wq->flush_cond);
284	}
285	update_accounting(sw);
286	pthread_cond_wait(&sw->cond, &sw->lock);
287	} else {
288	handle_work:
289	flist_splice_init(&sw->work_list, &local_list);
290	}
291	pthread_mutex_unlock(&sw->lock);
292	handle_list(sw, &local_list);
293	}
294
295	update_accounting(sw);
296
297	done:
298	pthread_mutex_lock(&sw->lock);
299	sw->flags \|= (SW_F_EXITED \| eflags);
300	pthread_mutex_unlock(&sw->lock);
301	return NULL;
302	}
303
304	static void free_worker(struct submit_worker *sw)
305	{
306	struct thread_data *td = &sw->td;
307
308	fio_options_free(td);
309	close_and_free_files(td);
310	if (td->io_ops)
311	close_ioengine(td);
312	td_set_runstate(td, TD_EXITED);
313
314	pthread_cond_destroy(&sw->cond);
315	pthread_mutex_destroy(&sw->lock);
316	}
317
318	static void shutdown_worker(struct submit_worker sw, unsigned int sum_cnt)
319	{
320	struct thread_data *parent = sw->wq->td;
321
322	pthread_join(sw->thread, NULL);
323	(*sum_cnt)++;
324	sum_thread_stats(&parent->ts, &sw->td.ts, *sum_cnt);
325	free_worker(sw);
326	}
327
328	void workqueue_exit(struct workqueue *wq)
329	{
330	unsigned int shutdown, sum_cnt = 0;
331	struct submit_worker *sw;
332	int i;
333
334	for (i = 0; i < wq->max_workers; i++) {
335	sw = &wq->workers[i];
336
337	pthread_mutex_lock(&sw->lock);
338	sw->flags \|= SW_F_EXIT;
339	pthread_cond_signal(&sw->cond);
340	pthread_mutex_unlock(&sw->lock);
341	}
342
343	do {
344	shutdown = 0;
345	for (i = 0; i < wq->max_workers; i++) {
346	sw = &wq->workers[i];
347	if (sw->flags & SW_F_ACCOUNTED)
348	continue;
349	sw->flags \|= SW_F_ACCOUNTED;
350	shutdown_worker(sw, &sum_cnt);
351	shutdown++;
352	}
353	} while (shutdown && shutdown != wq->max_workers);
354
355	free(wq->workers);
356	pthread_mutex_destroy(&wq->flush_lock);
357	pthread_cond_destroy(&wq->flush_cond);
358	}
359
360	static int start_worker(struct workqueue *wq, unsigned int index)
361	{
362	struct submit_worker *sw = &wq->workers[index];
363	int ret;
364
365	INIT_FLIST_HEAD(&sw->work_list);
366	pthread_cond_init(&sw->cond, NULL);
367	pthread_mutex_init(&sw->lock, NULL);
368	sw->wq = wq;
369	sw->index = index;
370
371	ret = pthread_create(&sw->thread, NULL, worker_thread, sw);
372	if (!ret) {
373	pthread_mutex_lock(&sw->lock);
374	sw->flags = SW_F_IDLE;
375	pthread_mutex_unlock(&sw->lock);
376	return 0;
377	}
378
379	free_worker(sw);
380	return 1;
381	}
382
383	int workqueue_init(struct thread_data td, struct workqueue wq,
384	workqueue_fn *fn, unsigned max_pending)
385	{
386	unsigned int running;
387	int i, error;
388
389	wq->max_workers = max_pending;
390	wq->td = td;
391	wq->fn = fn;
392	wq->work_seq = 0;
393	wq->next_free_worker = 0;
394	pthread_cond_init(&wq->flush_cond, NULL);
395	pthread_mutex_init(&wq->flush_lock, NULL);
396
397	wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker));
398
399	for (i = 0; i < wq->max_workers; i++)
400	if (start_worker(wq, i))
401	break;
402
403	wq->max_workers = i;
404	if (!wq->max_workers) {
405	err:
406	log_err("Can't create rate workqueue\n");
407	td_verror(td, ESRCH, "workqueue_init");
408	workqueue_exit(wq);
409	return 1;
410	}
411
412	/*
413	* Wait for them all to be started and initialized
414	*/
415	error = 0;
416	do {
417	struct submit_worker *sw;
418
419	running = 0;
420	pthread_mutex_lock(&wq->flush_lock);
421	for (i = 0; i < wq->max_workers; i++) {
422	sw = &wq->workers[i];
423	pthread_mutex_lock(&sw->lock);
424	if (sw->flags & SW_F_RUNNING)
425	running++;
426	if (sw->flags & SW_F_ERROR)
427	error++;
428	pthread_mutex_unlock(&sw->lock);
429	}
430
431	if (error \|\| running == wq->max_workers) {
432	pthread_mutex_unlock(&wq->flush_lock);
433	break;
434	}
435
436	pthread_cond_wait(&wq->flush_cond, &wq->flush_lock);
437	pthread_mutex_unlock(&wq->flush_lock);
438	} while (1);
439
440	if (error)
441	goto err;
442
443	return 0;
444	}