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" |
11 | #include "workqueue.h" | |
a9da8ab2 | 12 | |
a9da8ab2 JA |
13 | enum { |
14 | SW_F_IDLE = 1 << 0, | |
15 | SW_F_RUNNING = 1 << 1, | |
16 | SW_F_EXIT = 1 << 2, | |
17 | SW_F_EXITED = 1 << 3, | |
18 | SW_F_ACCOUNTED = 1 << 4, | |
19 | SW_F_ERROR = 1 << 5, | |
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 | } |
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 | |
122 | while (!flist_empty(list)) { | |
88271841 JA |
123 | work = flist_first_entry(list, struct workqueue_work, list); |
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; | |
f6496ba7 | 133 | unsigned int eflags = 0, ret = 0; |
a9da8ab2 JA |
134 | FLIST_HEAD(local_list); |
135 | ||
24660963 JA |
136 | sk_out_assign(sw->sk_out); |
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) |
205 | wq->ops.update_acct_fn(sw); | |
a9da8ab2 JA |
206 | |
207 | done: | |
208 | pthread_mutex_lock(&sw->lock); | |
209 | sw->flags |= (SW_F_EXITED | eflags); | |
210 | pthread_mutex_unlock(&sw->lock); | |
24660963 | 211 | sk_out_drop(); |
a9da8ab2 JA |
212 | return NULL; |
213 | } | |
214 | ||
c35c582d | 215 | static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt) |
a9da8ab2 | 216 | { |
ee2b6d6e | 217 | struct workqueue *wq = sw->wq; |
a9da8ab2 | 218 | |
c35c582d | 219 | workqueue_exit_worker(sw, sum_cnt); |
a9da8ab2 JA |
220 | |
221 | pthread_cond_destroy(&sw->cond); | |
222 | pthread_mutex_destroy(&sw->lock); | |
ee2b6d6e JA |
223 | |
224 | if (wq->ops.free_worker_fn) | |
225 | wq->ops.free_worker_fn(sw); | |
a9da8ab2 JA |
226 | } |
227 | ||
228 | static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt) | |
229 | { | |
a9da8ab2 | 230 | pthread_join(sw->thread, NULL); |
c35c582d | 231 | free_worker(sw, sum_cnt); |
a9da8ab2 JA |
232 | } |
233 | ||
234 | void workqueue_exit(struct workqueue *wq) | |
235 | { | |
236 | unsigned int shutdown, sum_cnt = 0; | |
237 | struct submit_worker *sw; | |
238 | int i; | |
239 | ||
f8e266c9 JA |
240 | if (!wq->workers) |
241 | return; | |
242 | ||
a9da8ab2 JA |
243 | for (i = 0; i < wq->max_workers; i++) { |
244 | sw = &wq->workers[i]; | |
245 | ||
246 | pthread_mutex_lock(&sw->lock); | |
247 | sw->flags |= SW_F_EXIT; | |
248 | pthread_cond_signal(&sw->cond); | |
249 | pthread_mutex_unlock(&sw->lock); | |
250 | } | |
251 | ||
252 | do { | |
253 | shutdown = 0; | |
254 | for (i = 0; i < wq->max_workers; i++) { | |
255 | sw = &wq->workers[i]; | |
256 | if (sw->flags & SW_F_ACCOUNTED) | |
257 | continue; | |
b7d0bbf1 | 258 | pthread_mutex_lock(&sw->lock); |
a9da8ab2 | 259 | sw->flags |= SW_F_ACCOUNTED; |
b7d0bbf1 | 260 | pthread_mutex_unlock(&sw->lock); |
a9da8ab2 JA |
261 | shutdown_worker(sw, &sum_cnt); |
262 | shutdown++; | |
263 | } | |
264 | } while (shutdown && shutdown != wq->max_workers); | |
265 | ||
266 | free(wq->workers); | |
f8e266c9 | 267 | wq->workers = NULL; |
a9da8ab2 JA |
268 | pthread_mutex_destroy(&wq->flush_lock); |
269 | pthread_cond_destroy(&wq->flush_cond); | |
2a274336 | 270 | pthread_mutex_destroy(&wq->stat_lock); |
a9da8ab2 JA |
271 | } |
272 | ||
24660963 JA |
273 | static int start_worker(struct workqueue *wq, unsigned int index, |
274 | struct sk_out *sk_out) | |
a9da8ab2 JA |
275 | { |
276 | struct submit_worker *sw = &wq->workers[index]; | |
277 | int ret; | |
278 | ||
279 | INIT_FLIST_HEAD(&sw->work_list); | |
280 | pthread_cond_init(&sw->cond, NULL); | |
281 | pthread_mutex_init(&sw->lock, NULL); | |
282 | sw->wq = wq; | |
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, |
306 | struct sk_out *sk_out) | |
a9da8ab2 JA |
307 | { |
308 | unsigned int running; | |
309 | int i, error; | |
310 | ||
24660963 | 311 | wq->max_workers = max_workers; |
a9da8ab2 | 312 | wq->td = td; |
5bb79f69 | 313 | wq->ops = *ops; |
a9da8ab2 JA |
314 | wq->work_seq = 0; |
315 | wq->next_free_worker = 0; | |
316 | pthread_cond_init(&wq->flush_cond, NULL); | |
317 | pthread_mutex_init(&wq->flush_lock, NULL); | |
2a274336 | 318 | pthread_mutex_init(&wq->stat_lock, NULL); |
a9da8ab2 JA |
319 | |
320 | wq->workers = calloc(wq->max_workers, sizeof(struct submit_worker)); | |
321 | ||
322 | for (i = 0; i < wq->max_workers; i++) | |
24660963 | 323 | if (start_worker(wq, i, sk_out)) |
a9da8ab2 JA |
324 | break; |
325 | ||
326 | wq->max_workers = i; | |
bddc8d16 JA |
327 | if (!wq->max_workers) |
328 | goto err; | |
a9da8ab2 JA |
329 | |
330 | /* | |
331 | * Wait for them all to be started and initialized | |
332 | */ | |
333 | error = 0; | |
334 | do { | |
335 | struct submit_worker *sw; | |
336 | ||
337 | running = 0; | |
338 | pthread_mutex_lock(&wq->flush_lock); | |
339 | for (i = 0; i < wq->max_workers; i++) { | |
340 | sw = &wq->workers[i]; | |
341 | pthread_mutex_lock(&sw->lock); | |
342 | if (sw->flags & SW_F_RUNNING) | |
343 | running++; | |
344 | if (sw->flags & SW_F_ERROR) | |
345 | error++; | |
346 | pthread_mutex_unlock(&sw->lock); | |
347 | } | |
348 | ||
349 | if (error || running == wq->max_workers) { | |
350 | pthread_mutex_unlock(&wq->flush_lock); | |
351 | break; | |
352 | } | |
353 | ||
354 | pthread_cond_wait(&wq->flush_cond, &wq->flush_lock); | |
355 | pthread_mutex_unlock(&wq->flush_lock); | |
356 | } while (1); | |
357 | ||
bddc8d16 JA |
358 | if (!error) |
359 | return 0; | |
a9da8ab2 | 360 | |
bddc8d16 JA |
361 | err: |
362 | log_err("Can't create rate workqueue\n"); | |
363 | td_verror(td, ESRCH, "workqueue_init"); | |
364 | workqueue_exit(wq); | |
365 | return 1; | |
a9da8ab2 | 366 | } |