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; | |
bddc8d16 JA |
404 | if (!wq->max_workers) |
405 | goto err; | |
a9da8ab2 JA |
406 | |
407 | /* | |
408 | * Wait for them all to be started and initialized | |
409 | */ | |
410 | error = 0; | |
411 | do { | |
412 | struct submit_worker *sw; | |
413 | ||
414 | running = 0; | |
415 | pthread_mutex_lock(&wq->flush_lock); | |
416 | for (i = 0; i < wq->max_workers; i++) { | |
417 | sw = &wq->workers[i]; | |
418 | pthread_mutex_lock(&sw->lock); | |
419 | if (sw->flags & SW_F_RUNNING) | |
420 | running++; | |
421 | if (sw->flags & SW_F_ERROR) | |
422 | error++; | |
423 | pthread_mutex_unlock(&sw->lock); | |
424 | } | |
425 | ||
426 | if (error || running == wq->max_workers) { | |
427 | pthread_mutex_unlock(&wq->flush_lock); | |
428 | break; | |
429 | } | |
430 | ||
431 | pthread_cond_wait(&wq->flush_cond, &wq->flush_lock); | |
432 | pthread_mutex_unlock(&wq->flush_lock); | |
433 | } while (1); | |
434 | ||
bddc8d16 JA |
435 | if (!error) |
436 | return 0; | |
a9da8ab2 | 437 | |
bddc8d16 JA |
438 | err: |
439 | log_err("Can't create rate workqueue\n"); | |
440 | td_verror(td, ESRCH, "workqueue_init"); | |
441 | workqueue_exit(wq); | |
442 | return 1; | |
a9da8ab2 | 443 | } |