• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Generic workqueue offload mechanism
3  *
4  * Copyright (C) 2015 Jens Axboe <axboe@kernel.dk>
5  *
6  */
7 #include <unistd.h>
8 
9 #include "fio.h"
10 #include "flist.h"
11 #include "workqueue.h"
12 #include "smalloc.h"
13 
14 enum {
15 	SW_F_IDLE	= 1 << 0,
16 	SW_F_RUNNING	= 1 << 1,
17 	SW_F_EXIT	= 1 << 2,
18 	SW_F_ACCOUNTED	= 1 << 3,
19 	SW_F_ERROR	= 1 << 4,
20 };
21 
__get_submit_worker(struct workqueue * wq,unsigned int start,unsigned int end,struct submit_worker ** best)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 
get_submit_worker(struct workqueue * wq)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 
all_sw_idle(struct workqueue * wq)68 static bool all_sw_idle(struct workqueue *wq)
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))
76 			return false;
77 	}
78 
79 	return true;
80 }
81 
82 /*
83  * Must be serialized wrt workqueue_enqueue() by caller
84  */
workqueue_flush(struct workqueue * wq)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 /*
99  * Must be serialized by caller. Returns true for queued, false for busy.
100  */
workqueue_enqueue(struct workqueue * wq,struct workqueue_work * work)101 void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work)
102 {
103 	struct submit_worker *sw;
104 
105 	sw = get_submit_worker(wq);
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);
113 
114 	pthread_cond_signal(&sw->cond);
115 }
116 
handle_list(struct submit_worker * sw,struct flist_head * list)117 static void handle_list(struct submit_worker *sw, struct flist_head *list)
118 {
119 	struct workqueue *wq = sw->wq;
120 	struct workqueue_work *work;
121 
122 	while (!flist_empty(list)) {
123 		work = flist_first_entry(list, struct workqueue_work, list);
124 		flist_del_init(&work->list);
125 		wq->ops.fn(sw, work);
126 	}
127 }
128 
worker_thread(void * data)129 static void *worker_thread(void *data)
130 {
131 	struct submit_worker *sw = data;
132 	struct workqueue *wq = sw->wq;
133 	unsigned int ret = 0;
134 	FLIST_HEAD(local_list);
135 
136 	sk_out_assign(sw->sk_out);
137 
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 
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 
170 			if (workqueue_pre_sleep_check(sw)) {
171 				pthread_mutex_unlock(&sw->lock);
172 				workqueue_pre_sleep(sw);
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 			}
192 			if (wq->ops.update_acct_fn)
193 				wq->ops.update_acct_fn(sw);
194 
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 
204 	if (wq->ops.update_acct_fn)
205 		wq->ops.update_acct_fn(sw);
206 
207 done:
208 	sk_out_drop();
209 	return NULL;
210 }
211 
free_worker(struct submit_worker * sw,unsigned int * sum_cnt)212 static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt)
213 {
214 	struct workqueue *wq = sw->wq;
215 
216 	workqueue_exit_worker(sw, sum_cnt);
217 
218 	pthread_cond_destroy(&sw->cond);
219 	pthread_mutex_destroy(&sw->lock);
220 
221 	if (wq->ops.free_worker_fn)
222 		wq->ops.free_worker_fn(sw);
223 }
224 
shutdown_worker(struct submit_worker * sw,unsigned int * sum_cnt)225 static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt)
226 {
227 	pthread_join(sw->thread, NULL);
228 	free_worker(sw, sum_cnt);
229 }
230 
workqueue_exit(struct workqueue * wq)231 void workqueue_exit(struct workqueue *wq)
232 {
233 	unsigned int shutdown, sum_cnt = 0;
234 	struct submit_worker *sw;
235 	int i;
236 
237 	if (!wq->workers)
238 		return;
239 
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;
255 			pthread_mutex_lock(&sw->lock);
256 			sw->flags |= SW_F_ACCOUNTED;
257 			pthread_mutex_unlock(&sw->lock);
258 			shutdown_worker(sw, &sum_cnt);
259 			shutdown++;
260 		}
261 	} while (shutdown && shutdown != wq->max_workers);
262 
263 	sfree(wq->workers);
264 	wq->workers = NULL;
265 	pthread_mutex_destroy(&wq->flush_lock);
266 	pthread_cond_destroy(&wq->flush_cond);
267 	pthread_mutex_destroy(&wq->stat_lock);
268 }
269 
start_worker(struct workqueue * wq,unsigned int index,struct sk_out * sk_out)270 static int start_worker(struct workqueue *wq, unsigned int index,
271 			struct sk_out *sk_out)
272 {
273 	struct submit_worker *sw = &wq->workers[index];
274 	int ret;
275 
276 	INIT_FLIST_HEAD(&sw->work_list);
277 
278 	ret = mutex_cond_init_pshared(&sw->lock, &sw->cond);
279 	if (ret)
280 		return ret;
281 
282 	sw->wq = wq;
283 	sw->index = index;
284 	sw->sk_out = sk_out;
285 
286 	if (wq->ops.alloc_worker_fn) {
287 		ret = wq->ops.alloc_worker_fn(sw);
288 		if (ret)
289 			return ret;
290 	}
291 
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 
300 	free_worker(sw, NULL);
301 	return 1;
302 }
303 
workqueue_init(struct thread_data * td,struct workqueue * wq,struct workqueue_ops * ops,unsigned int max_workers,struct sk_out * sk_out)304 int workqueue_init(struct thread_data *td, struct workqueue *wq,
305 		   struct workqueue_ops *ops, unsigned int max_workers,
306 		   struct sk_out *sk_out)
307 {
308 	unsigned int running;
309 	int i, error;
310 	int ret;
311 
312 	wq->max_workers = max_workers;
313 	wq->td = td;
314 	wq->ops = *ops;
315 	wq->work_seq = 0;
316 	wq->next_free_worker = 0;
317 
318 	ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond);
319 	if (ret)
320 		goto err;
321 	ret = mutex_init_pshared(&wq->stat_lock);
322 	if (ret)
323 		goto err;
324 
325 	wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker));
326 	if (!wq->workers)
327 		goto err;
328 
329 	for (i = 0; i < wq->max_workers; i++)
330 		if (start_worker(wq, i, sk_out))
331 			break;
332 
333 	wq->max_workers = i;
334 	if (!wq->max_workers)
335 		goto err;
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 
365 	if (!error)
366 		return 0;
367 
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;
373 }
374