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