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" | |
a47591e4 | 12 | #include "smalloc.h" |
ae626d4e | 13 | #include "pshared.h" |
a9da8ab2 | 14 | |
a9da8ab2 JA |
15 | enum { |
16 | SW_F_IDLE = 1 << 0, | |
17 | SW_F_RUNNING = 1 << 1, | |
18 | SW_F_EXIT = 1 << 2, | |
d22042d2 JA |
19 | SW_F_ACCOUNTED = 1 << 3, |
20 | SW_F_ERROR = 1 << 4, | |
a9da8ab2 JA |
21 | }; |
22 | ||
23 | static struct submit_worker *__get_submit_worker(struct workqueue *wq, | |
24 | unsigned int start, | |
25 | unsigned int end, | |
26 | struct submit_worker **best) | |
27 | { | |
28 | struct submit_worker *sw = NULL; | |
29 | ||
30 | while (start <= end) { | |
31 | sw = &wq->workers[start]; | |
32 | if (sw->flags & SW_F_IDLE) | |
33 | return sw; | |
34 | if (!(*best) || sw->seq < (*best)->seq) | |
35 | *best = sw; | |
36 | start++; | |
37 | } | |
38 | ||
39 | return NULL; | |
40 | } | |
41 | ||
42 | static struct submit_worker *get_submit_worker(struct workqueue *wq) | |
43 | { | |
44 | unsigned int next = wq->next_free_worker; | |
45 | struct submit_worker *sw, *best = NULL; | |
46 | ||
47 | assert(next < wq->max_workers); | |
48 | ||
49 | sw = __get_submit_worker(wq, next, wq->max_workers - 1, &best); | |
50 | if (!sw && next) | |
51 | sw = __get_submit_worker(wq, 0, next - 1, &best); | |
52 | ||
53 | /* | |
54 | * No truly idle found, use best match | |
55 | */ | |
56 | if (!sw) | |
57 | sw = best; | |
58 | ||
59 | if (sw->index == wq->next_free_worker) { | |
60 | if (sw->index + 1 < wq->max_workers) | |
61 | wq->next_free_worker = sw->index + 1; | |
62 | else | |
63 | wq->next_free_worker = 0; | |
64 | } | |
65 | ||
66 | return sw; | |
67 | } | |
68 | ||
1391052a | 69 | static bool all_sw_idle(struct workqueue *wq) |
a9da8ab2 JA |
70 | { |
71 | int i; | |
72 | ||
73 | for (i = 0; i < wq->max_workers; i++) { | |
74 | struct submit_worker *sw = &wq->workers[i]; | |
75 | ||
76 | if (!(sw->flags & SW_F_IDLE)) | |
1391052a | 77 | return false; |
a9da8ab2 JA |
78 | } |
79 | ||
1391052a | 80 | return true; |
a9da8ab2 JA |
81 | } |
82 | ||
83 | /* | |
84 | * Must be serialized wrt workqueue_enqueue() by caller | |
85 | */ | |
86 | void workqueue_flush(struct workqueue *wq) | |
87 | { | |
88 | wq->wake_idle = 1; | |
89 | ||
90 | while (!all_sw_idle(wq)) { | |
91 | pthread_mutex_lock(&wq->flush_lock); | |
92 | pthread_cond_wait(&wq->flush_cond, &wq->flush_lock); | |
93 | pthread_mutex_unlock(&wq->flush_lock); | |
94 | } | |
95 | ||
96 | wq->wake_idle = 0; | |
97 | } | |
98 | ||
99 | /* | |
b07f6ad1 | 100 | * Must be serialized by caller. Returns true for queued, false for busy. |
a9da8ab2 | 101 | */ |
26de50cf | 102 | void workqueue_enqueue(struct workqueue *wq, struct workqueue_work *work) |
a9da8ab2 JA |
103 | { |
104 | struct submit_worker *sw; | |
105 | ||
106 | sw = get_submit_worker(wq); | |
26de50cf JA |
107 | assert(sw); |
108 | ||
109 | pthread_mutex_lock(&sw->lock); | |
110 | flist_add_tail(&work->list, &sw->work_list); | |
111 | sw->seq = ++wq->work_seq; | |
112 | sw->flags &= ~SW_F_IDLE; | |
a9da8ab2 | 113 | |
26de50cf | 114 | pthread_cond_signal(&sw->cond); |
b93a8eb0 | 115 | pthread_mutex_unlock(&sw->lock); |
a9da8ab2 JA |
116 | } |
117 | ||
118 | static void handle_list(struct submit_worker *sw, struct flist_head *list) | |
119 | { | |
120 | struct workqueue *wq = sw->wq; | |
88271841 | 121 | struct workqueue_work *work; |
a9da8ab2 JA |
122 | |
123 | while (!flist_empty(list)) { | |
88271841 JA |
124 | work = flist_first_entry(list, struct workqueue_work, list); |
125 | flist_del_init(&work->list); | |
ee2b6d6e | 126 | wq->ops.fn(sw, work); |
a9da8ab2 JA |
127 | } |
128 | } | |
129 | ||
a9da8ab2 JA |
130 | static void *worker_thread(void *data) |
131 | { | |
132 | struct submit_worker *sw = data; | |
133 | struct workqueue *wq = sw->wq; | |
d22042d2 | 134 | unsigned int ret = 0; |
a9da8ab2 JA |
135 | FLIST_HEAD(local_list); |
136 | ||
24660963 JA |
137 | sk_out_assign(sw->sk_out); |
138 | ||
f6496ba7 JA |
139 | if (wq->ops.nice) { |
140 | if (nice(wq->ops.nice) < 0) { | |
141 | log_err("workqueue: nice %s\n", strerror(errno)); | |
142 | ret = 1; | |
143 | } | |
144 | } | |
145 | ||
146 | if (!ret) | |
147 | ret = workqueue_init_worker(sw); | |
148 | ||
a9da8ab2 JA |
149 | pthread_mutex_lock(&sw->lock); |
150 | sw->flags |= SW_F_RUNNING; | |
151 | if (ret) | |
152 | sw->flags |= SW_F_ERROR; | |
153 | pthread_mutex_unlock(&sw->lock); | |
154 | ||
155 | pthread_mutex_lock(&wq->flush_lock); | |
156 | pthread_cond_signal(&wq->flush_cond); | |
157 | pthread_mutex_unlock(&wq->flush_lock); | |
158 | ||
159 | if (sw->flags & SW_F_ERROR) | |
160 | goto done; | |
161 | ||
162 | while (1) { | |
163 | pthread_mutex_lock(&sw->lock); | |
164 | ||
165 | if (flist_empty(&sw->work_list)) { | |
166 | if (sw->flags & SW_F_EXIT) { | |
167 | pthread_mutex_unlock(&sw->lock); | |
168 | break; | |
169 | } | |
170 | ||
ee2b6d6e | 171 | if (workqueue_pre_sleep_check(sw)) { |
a9da8ab2 | 172 | pthread_mutex_unlock(&sw->lock); |
ee2b6d6e | 173 | workqueue_pre_sleep(sw); |
a9da8ab2 JA |
174 | pthread_mutex_lock(&sw->lock); |
175 | } | |
176 | ||
177 | /* | |
178 | * We dropped and reaquired the lock, check | |
179 | * state again. | |
180 | */ | |
181 | if (!flist_empty(&sw->work_list)) | |
182 | goto handle_work; | |
183 | ||
184 | if (sw->flags & SW_F_EXIT) { | |
185 | pthread_mutex_unlock(&sw->lock); | |
186 | break; | |
187 | } else if (!(sw->flags & SW_F_IDLE)) { | |
188 | sw->flags |= SW_F_IDLE; | |
189 | wq->next_free_worker = sw->index; | |
190 | if (wq->wake_idle) | |
191 | pthread_cond_signal(&wq->flush_cond); | |
192 | } | |
17ecadc3 JA |
193 | if (wq->ops.update_acct_fn) |
194 | wq->ops.update_acct_fn(sw); | |
195 | ||
a9da8ab2 JA |
196 | pthread_cond_wait(&sw->cond, &sw->lock); |
197 | } else { | |
198 | handle_work: | |
199 | flist_splice_init(&sw->work_list, &local_list); | |
200 | } | |
201 | pthread_mutex_unlock(&sw->lock); | |
202 | handle_list(sw, &local_list); | |
203 | } | |
204 | ||
17ecadc3 JA |
205 | if (wq->ops.update_acct_fn) |
206 | wq->ops.update_acct_fn(sw); | |
a9da8ab2 JA |
207 | |
208 | done: | |
24660963 | 209 | sk_out_drop(); |
a9da8ab2 JA |
210 | return NULL; |
211 | } | |
212 | ||
c35c582d | 213 | static void free_worker(struct submit_worker *sw, unsigned int *sum_cnt) |
a9da8ab2 | 214 | { |
ee2b6d6e | 215 | struct workqueue *wq = sw->wq; |
a9da8ab2 | 216 | |
c35c582d | 217 | workqueue_exit_worker(sw, sum_cnt); |
a9da8ab2 JA |
218 | |
219 | pthread_cond_destroy(&sw->cond); | |
220 | pthread_mutex_destroy(&sw->lock); | |
ee2b6d6e JA |
221 | |
222 | if (wq->ops.free_worker_fn) | |
223 | wq->ops.free_worker_fn(sw); | |
a9da8ab2 JA |
224 | } |
225 | ||
226 | static void shutdown_worker(struct submit_worker *sw, unsigned int *sum_cnt) | |
227 | { | |
a9da8ab2 | 228 | pthread_join(sw->thread, NULL); |
c35c582d | 229 | free_worker(sw, sum_cnt); |
a9da8ab2 JA |
230 | } |
231 | ||
232 | void workqueue_exit(struct workqueue *wq) | |
233 | { | |
234 | unsigned int shutdown, sum_cnt = 0; | |
235 | struct submit_worker *sw; | |
236 | int i; | |
237 | ||
f8e266c9 JA |
238 | if (!wq->workers) |
239 | return; | |
240 | ||
a9da8ab2 JA |
241 | for (i = 0; i < wq->max_workers; i++) { |
242 | sw = &wq->workers[i]; | |
243 | ||
244 | pthread_mutex_lock(&sw->lock); | |
245 | sw->flags |= SW_F_EXIT; | |
246 | pthread_cond_signal(&sw->cond); | |
247 | pthread_mutex_unlock(&sw->lock); | |
248 | } | |
249 | ||
250 | do { | |
251 | shutdown = 0; | |
252 | for (i = 0; i < wq->max_workers; i++) { | |
253 | sw = &wq->workers[i]; | |
254 | if (sw->flags & SW_F_ACCOUNTED) | |
255 | continue; | |
b7d0bbf1 | 256 | pthread_mutex_lock(&sw->lock); |
a9da8ab2 | 257 | sw->flags |= SW_F_ACCOUNTED; |
b7d0bbf1 | 258 | pthread_mutex_unlock(&sw->lock); |
a9da8ab2 JA |
259 | shutdown_worker(sw, &sum_cnt); |
260 | shutdown++; | |
261 | } | |
262 | } while (shutdown && shutdown != wq->max_workers); | |
263 | ||
a47591e4 | 264 | sfree(wq->workers); |
f8e266c9 | 265 | wq->workers = NULL; |
a9da8ab2 JA |
266 | pthread_mutex_destroy(&wq->flush_lock); |
267 | pthread_cond_destroy(&wq->flush_cond); | |
2a274336 | 268 | pthread_mutex_destroy(&wq->stat_lock); |
a9da8ab2 JA |
269 | } |
270 | ||
24660963 JA |
271 | static int start_worker(struct workqueue *wq, unsigned int index, |
272 | struct sk_out *sk_out) | |
a9da8ab2 JA |
273 | { |
274 | struct submit_worker *sw = &wq->workers[index]; | |
275 | int ret; | |
276 | ||
277 | INIT_FLIST_HEAD(&sw->work_list); | |
34febb23 JA |
278 | |
279 | ret = mutex_cond_init_pshared(&sw->lock, &sw->cond); | |
f9e5b5ee JK |
280 | if (ret) |
281 | return ret; | |
34febb23 | 282 | |
a9da8ab2 JA |
283 | sw->wq = wq; |
284 | sw->index = index; | |
24660963 | 285 | sw->sk_out = sk_out; |
a9da8ab2 | 286 | |
ee2b6d6e JA |
287 | if (wq->ops.alloc_worker_fn) { |
288 | ret = wq->ops.alloc_worker_fn(sw); | |
289 | if (ret) | |
290 | return ret; | |
291 | } | |
292 | ||
a9da8ab2 JA |
293 | ret = pthread_create(&sw->thread, NULL, worker_thread, sw); |
294 | if (!ret) { | |
295 | pthread_mutex_lock(&sw->lock); | |
296 | sw->flags = SW_F_IDLE; | |
297 | pthread_mutex_unlock(&sw->lock); | |
298 | return 0; | |
299 | } | |
300 | ||
c35c582d | 301 | free_worker(sw, NULL); |
a9da8ab2 JA |
302 | return 1; |
303 | } | |
304 | ||
305 | int workqueue_init(struct thread_data *td, struct workqueue *wq, | |
24660963 JA |
306 | struct workqueue_ops *ops, unsigned int max_workers, |
307 | struct sk_out *sk_out) | |
a9da8ab2 JA |
308 | { |
309 | unsigned int running; | |
310 | int i, error; | |
f9e5b5ee | 311 | int ret; |
a9da8ab2 | 312 | |
24660963 | 313 | wq->max_workers = max_workers; |
a9da8ab2 | 314 | wq->td = td; |
5bb79f69 | 315 | wq->ops = *ops; |
a9da8ab2 JA |
316 | wq->work_seq = 0; |
317 | wq->next_free_worker = 0; | |
f9e5b5ee | 318 | |
34febb23 JA |
319 | ret = mutex_cond_init_pshared(&wq->flush_lock, &wq->flush_cond); |
320 | if (ret) | |
f9e5b5ee | 321 | goto err; |
34febb23 JA |
322 | ret = mutex_init_pshared(&wq->stat_lock); |
323 | if (ret) | |
f9e5b5ee | 324 | goto err; |
a9da8ab2 | 325 | |
a47591e4 | 326 | wq->workers = smalloc(wq->max_workers * sizeof(struct submit_worker)); |
81b3c86f JA |
327 | if (!wq->workers) |
328 | goto err; | |
a9da8ab2 JA |
329 | |
330 | for (i = 0; i < wq->max_workers; i++) | |
24660963 | 331 | if (start_worker(wq, i, sk_out)) |
a9da8ab2 JA |
332 | break; |
333 | ||
334 | wq->max_workers = i; | |
bddc8d16 JA |
335 | if (!wq->max_workers) |
336 | goto err; | |
a9da8ab2 JA |
337 | |
338 | /* | |
339 | * Wait for them all to be started and initialized | |
340 | */ | |
341 | error = 0; | |
342 | do { | |
343 | struct submit_worker *sw; | |
344 | ||
345 | running = 0; | |
346 | pthread_mutex_lock(&wq->flush_lock); | |
347 | for (i = 0; i < wq->max_workers; i++) { | |
348 | sw = &wq->workers[i]; | |
349 | pthread_mutex_lock(&sw->lock); | |
350 | if (sw->flags & SW_F_RUNNING) | |
351 | running++; | |
352 | if (sw->flags & SW_F_ERROR) | |
353 | error++; | |
354 | pthread_mutex_unlock(&sw->lock); | |
355 | } | |
356 | ||
357 | if (error || running == wq->max_workers) { | |
358 | pthread_mutex_unlock(&wq->flush_lock); | |
359 | break; | |
360 | } | |
361 | ||
362 | pthread_cond_wait(&wq->flush_cond, &wq->flush_lock); | |
363 | pthread_mutex_unlock(&wq->flush_lock); | |
364 | } while (1); | |
365 | ||
bddc8d16 JA |
366 | if (!error) |
367 | return 0; | |
a9da8ab2 | 368 | |
bddc8d16 JA |
369 | err: |
370 | log_err("Can't create rate workqueue\n"); | |
371 | td_verror(td, ESRCH, "workqueue_init"); | |
372 | workqueue_exit(wq); | |
373 | return 1; | |
a9da8ab2 | 374 | } |