Commit | Line | Data |
---|---|---|
f2a2ce0e HL |
1 | #include <math.h> |
2 | #include "json.h" | |
3 | #include "idletime.h" | |
4 | ||
5 | static volatile struct idle_prof_common ipc; | |
6 | ||
7 | /* Get time to complete an unit work on a particular cpu. | |
8 | * The minimum number in CALIBRATE_RUNS runs is returned. | |
9 | */ | |
10 | static double calibrate_unit(unsigned char *data) | |
11 | { | |
12 | unsigned long t, i, j, k; | |
13 | struct timeval tps; | |
14 | double tunit = 0.0; | |
15 | ||
16 | for (i=0; i<CALIBRATE_RUNS; i++) { | |
17 | ||
18 | fio_gettime(&tps, NULL); | |
19 | /* scale for less variance */ | |
20 | for (j=0; j < CALIBRATE_SCALE; j++) { | |
21 | /* unit of work */ | |
22 | for (k=0; k < page_size; k++) { | |
23 | data[(k+j)%page_size] = k%256; | |
24 | /* we won't see STOP here. this is to match | |
25 | * the same statement in the profiling loop. | |
26 | */ | |
27 | if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) | |
28 | return 0.0; | |
29 | } | |
30 | } | |
31 | ||
32 | t = utime_since_now(&tps); | |
33 | if (!t) | |
34 | continue; | |
35 | ||
36 | /* get the minimum time to complete CALIBRATE_SCALE units */ | |
37 | if ((i==0) || ((double)t < tunit)) | |
38 | tunit = (double)t; | |
39 | } | |
40 | ||
41 | return tunit/CALIBRATE_SCALE; | |
42 | } | |
43 | ||
44 | static void *idle_prof_thread_fn(void *data) | |
45 | { | |
46 | int retval; | |
47 | unsigned long j, k; | |
48 | struct idle_prof_thread *ipt = data; | |
49 | ||
50 | /* wait for all threads are spawned */ | |
51 | pthread_mutex_lock(&ipt->init_lock); | |
52 | ||
53 | /* exit if any other thread failed to start */ | |
54 | if (ipc.status == IDLE_PROF_STATUS_ABORT) | |
55 | return NULL; | |
56 | ||
57 | #if defined(FIO_HAVE_CPU_AFFINITY) | |
58 | os_cpu_mask_t cpu_mask; | |
59 | memset(&cpu_mask, 0, sizeof(cpu_mask)); | |
60 | fio_cpu_set(&cpu_mask, ipt->cpu); | |
61 | ||
62 | if ((retval=fio_setaffinity(gettid(), cpu_mask)) == -1) | |
63 | log_err("fio: fio_setaffinity failed\n"); | |
64 | #else | |
65 | retval = -1; | |
66 | log_err("fio: fio_setaffinity not supported\n"); | |
67 | #endif | |
68 | if (retval == -1) { | |
69 | ipt->state = TD_EXITED; | |
70 | pthread_mutex_unlock(&ipt->init_lock); | |
71 | return NULL; | |
72 | } | |
73 | ||
74 | ipt->cali_time = calibrate_unit(ipt->data); | |
75 | ||
76 | /* delay to set IDLE class till now for better calibration accuracy */ | |
77 | #if defined(FIO_HAVE_SCHED_IDLE) | |
78 | if ((retval = fio_set_sched_idle())) | |
79 | log_err("fio: fio_set_sched_idle failed\n"); | |
80 | #else | |
81 | retval = -1; | |
82 | log_err("fio: fio_set_sched_idle not supported\n"); | |
83 | #endif | |
84 | if (retval == -1) { | |
85 | ipt->state = TD_EXITED; | |
86 | pthread_mutex_unlock(&ipt->init_lock); | |
87 | return NULL; | |
88 | } | |
89 | ||
90 | ipt->state = TD_INITIALIZED; | |
91 | ||
92 | /* signal the main thread that calibration is done */ | |
93 | pthread_cond_signal(&ipt->cond); | |
94 | pthread_mutex_unlock(&ipt->init_lock); | |
95 | ||
96 | /* wait for other calibration to finish */ | |
97 | pthread_mutex_lock(&ipt->start_lock); | |
98 | ||
99 | /* exit if other threads failed to initialize */ | |
100 | if (ipc.status == IDLE_PROF_STATUS_ABORT) | |
101 | return NULL; | |
102 | ||
103 | /* exit if we are doing calibration only */ | |
104 | if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) | |
105 | return NULL; | |
106 | ||
107 | fio_gettime(&ipt->tps, NULL); | |
108 | ipt->state = TD_RUNNING; | |
109 | ||
110 | j = 0; | |
111 | while (1) { | |
112 | for (k=0; k < page_size; k++) { | |
113 | ipt->data[(k+j)%page_size] = k%256; | |
114 | if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) { | |
115 | fio_gettime(&ipt->tpe, NULL); | |
116 | goto idle_prof_done; | |
117 | } | |
118 | } | |
119 | j++; | |
120 | } | |
121 | ||
122 | idle_prof_done: | |
123 | ||
124 | ipt->loops = j + (double)k/page_size; | |
125 | ipt->state = TD_EXITED; | |
126 | pthread_mutex_unlock(&ipt->start_lock); | |
127 | ||
128 | return NULL; | |
129 | } | |
130 | ||
131 | /* calculate mean and standard deviation to complete an unit of work */ | |
132 | static void calibration_stats(void) | |
133 | { | |
134 | int i; | |
135 | double sum=0.0, var=0.0; | |
136 | struct idle_prof_thread *ipt; | |
137 | ||
138 | for (i = 0; i < ipc.nr_cpus; i++) { | |
139 | ipt = &ipc.ipts[i]; | |
140 | sum += ipt->cali_time; | |
141 | } | |
142 | ||
143 | ipc.cali_mean = sum/ipc.nr_cpus; | |
144 | ||
145 | for (i = 0; i < ipc.nr_cpus; i++) { | |
146 | ipt = &ipc.ipts[i]; | |
147 | var += pow(ipt->cali_time-ipc.cali_mean, 2); | |
148 | } | |
149 | ||
150 | ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1)); | |
151 | } | |
152 | ||
153 | void fio_idle_prof_init(void) | |
154 | { | |
155 | int i, ret; | |
156 | struct timeval tp; | |
157 | struct timespec ts; | |
158 | pthread_attr_t tattr; | |
159 | struct idle_prof_thread *ipt; | |
160 | ||
161 | ipc.nr_cpus = cpus_online(); | |
162 | ipc.status = IDLE_PROF_STATUS_OK; | |
163 | ||
164 | if (ipc.opt == IDLE_PROF_OPT_NONE) | |
165 | return; | |
166 | ||
167 | if ((ret = pthread_attr_init(&tattr))) { | |
168 | log_err("fio: pthread_attr_init %s\n", strerror(ret)); | |
169 | return; | |
170 | } | |
171 | if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) { | |
172 | log_err("fio: pthread_attr_setscope %s\n", strerror(ret)); | |
173 | return; | |
174 | } | |
175 | ||
176 | ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread)); | |
177 | if (!ipc.ipts) { | |
178 | log_err("fio: malloc failed\n"); | |
179 | return; | |
180 | } | |
181 | ||
182 | ipc.buf = malloc(ipc.nr_cpus * page_size); | |
183 | if (!ipc.buf) { | |
184 | log_err("fio: malloc failed\n"); | |
185 | free(ipc.ipts); | |
186 | return; | |
187 | } | |
188 | ||
189 | /* profiling aborts on any single thread failure since the | |
190 | * result won't be accurate if any cpu is not used. | |
191 | */ | |
192 | for (i = 0; i < ipc.nr_cpus; i++) { | |
193 | ipt = &ipc.ipts[i]; | |
194 | ||
195 | ipt->cpu = i; | |
196 | ipt->state = TD_NOT_CREATED; | |
197 | ipt->data = (unsigned char *)(ipc.buf + page_size * i); | |
198 | ||
199 | if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) { | |
200 | ipc.status = IDLE_PROF_STATUS_ABORT; | |
201 | log_err("fio: pthread_mutex_init %s\n", strerror(ret)); | |
202 | break; | |
203 | } | |
204 | ||
205 | if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) { | |
206 | ipc.status = IDLE_PROF_STATUS_ABORT; | |
207 | log_err("fio: pthread_mutex_init %s\n", strerror(ret)); | |
208 | break; | |
209 | } | |
210 | ||
211 | if ((ret = pthread_cond_init(&ipt->cond, NULL))) { | |
212 | ipc.status = IDLE_PROF_STATUS_ABORT; | |
213 | log_err("fio: pthread_cond_init %s\n", strerror(ret)); | |
214 | break; | |
215 | } | |
216 | ||
217 | /* make sure all threads are spawned before they start */ | |
218 | pthread_mutex_lock(&ipt->init_lock); | |
219 | ||
220 | /* make sure all threads finish init before profiling starts */ | |
221 | pthread_mutex_lock(&ipt->start_lock); | |
222 | ||
223 | if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) { | |
224 | ipc.status = IDLE_PROF_STATUS_ABORT; | |
225 | log_err("fio: pthread_create %s\n", strerror(ret)); | |
226 | break; | |
227 | } else { | |
228 | ipt->state = TD_CREATED; | |
229 | } | |
230 | ||
231 | if ((ret = pthread_detach(ipt->thread))) { | |
232 | /* log error and let the thread spin */ | |
233 | log_err("fio: pthread_detatch %s\n", strerror(ret)); | |
234 | } | |
235 | } | |
236 | ||
237 | /* let good threads continue so that they can exit | |
238 | * if errors on other threads occurred previously. | |
239 | */ | |
240 | for (i = 0; i < ipc.nr_cpus; i++) { | |
241 | ipt = &ipc.ipts[i]; | |
242 | pthread_mutex_unlock(&ipt->init_lock); | |
243 | } | |
244 | ||
245 | if (ipc.status == IDLE_PROF_STATUS_ABORT) | |
246 | return; | |
247 | ||
248 | /* wait for calibration to finish */ | |
249 | for (i = 0; i < ipc.nr_cpus; i++) { | |
250 | ipt = &ipc.ipts[i]; | |
251 | pthread_mutex_lock(&ipt->init_lock); | |
252 | while ((ipt->state!=TD_EXITED) && (ipt->state!=TD_INITIALIZED)) { | |
253 | fio_gettime(&tp, NULL); | |
254 | ts.tv_sec = tp.tv_sec + 1; | |
255 | ts.tv_nsec = tp.tv_usec * 1000; | |
256 | pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts); | |
257 | } | |
258 | pthread_mutex_unlock(&ipt->init_lock); | |
259 | ||
260 | /* any thread failed to initialize would abort other threads | |
261 | * later after fio_idle_prof_start. | |
262 | */ | |
263 | if (ipt->state == TD_EXITED) | |
264 | ipc.status = IDLE_PROF_STATUS_ABORT; | |
265 | } | |
266 | ||
267 | if (ipc.status != IDLE_PROF_STATUS_ABORT) | |
268 | calibration_stats(); | |
269 | else | |
270 | ipc.cali_mean = ipc.cali_stddev = 0.0; | |
271 | ||
272 | if (ipc.opt == IDLE_PROF_OPT_CALI) | |
273 | ipc.status = IDLE_PROF_STATUS_CALI_STOP; | |
274 | } | |
275 | ||
276 | void fio_idle_prof_start(void) | |
277 | { | |
278 | int i; | |
279 | struct idle_prof_thread *ipt; | |
280 | ||
281 | if (ipc.opt == IDLE_PROF_OPT_NONE) | |
282 | return; | |
283 | ||
284 | /* unlock regardless abort is set or not */ | |
285 | for (i = 0; i < ipc.nr_cpus; i++) { | |
286 | ipt = &ipc.ipts[i]; | |
287 | pthread_mutex_unlock(&ipt->start_lock); | |
288 | } | |
289 | } | |
290 | ||
291 | void fio_idle_prof_stop(void) | |
292 | { | |
293 | int i; | |
294 | uint64_t runt; | |
295 | struct timeval tp; | |
296 | struct timespec ts; | |
297 | struct idle_prof_thread *ipt; | |
298 | ||
299 | if (ipc.opt == IDLE_PROF_OPT_NONE) | |
300 | return; | |
301 | ||
302 | if (ipc.opt == IDLE_PROF_OPT_CALI) | |
303 | return; | |
304 | ||
305 | ipc.status = IDLE_PROF_STATUS_PROF_STOP; | |
306 | ||
307 | /* wait for all threads to exit from profiling */ | |
308 | for (i = 0; i < ipc.nr_cpus; i++) { | |
309 | ipt = &ipc.ipts[i]; | |
310 | pthread_mutex_lock(&ipt->start_lock); | |
311 | while ((ipt->state!=TD_EXITED) && (ipt->state!=TD_NOT_CREATED)) { | |
312 | fio_gettime(&tp, NULL); | |
313 | ts.tv_sec = tp.tv_sec + 1; | |
314 | ts.tv_nsec = tp.tv_usec * 1000; | |
315 | /* timed wait in case a signal is not received */ | |
316 | pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts); | |
317 | } | |
318 | pthread_mutex_unlock(&ipt->start_lock); | |
319 | ||
320 | /* calculate idleness */ | |
321 | if (ipc.cali_mean != 0.0) { | |
322 | runt = utime_since(&ipt->tps, &ipt->tpe); | |
323 | ipt->idleness = ipt->loops * ipc.cali_mean / runt; | |
324 | } else | |
325 | ipt->idleness = 0.0; | |
326 | } | |
327 | ||
328 | /* memory allocations are freed via explicit fio_idle_prof_cleanup | |
329 | * after profiling stats are collected by apps. | |
330 | */ | |
331 | ||
332 | return; | |
333 | } | |
334 | ||
335 | /* return system idle percentage when cpu is -1; | |
336 | * return one cpu idle percentage otherwise. | |
337 | */ | |
338 | static double fio_idle_prof_cpu_stat(int cpu) | |
339 | { | |
340 | int i, nr_cpus = ipc.nr_cpus; | |
341 | struct idle_prof_thread *ipt; | |
342 | double p = 0.0; | |
343 | ||
344 | if (ipc.opt == IDLE_PROF_OPT_NONE) | |
345 | return 0.0; | |
346 | ||
347 | if ((cpu >= nr_cpus) || (cpu < -1)) { | |
348 | log_err("fio: idle profiling invalid cpu index\n"); | |
349 | return 0.0; | |
350 | } | |
351 | ||
352 | if (cpu == -1) { | |
353 | for (i = 0; i < nr_cpus; i++) { | |
354 | ipt = &ipc.ipts[i]; | |
355 | p += ipt->idleness; | |
356 | } | |
357 | p /= nr_cpus; | |
358 | } else { | |
359 | ipt = &ipc.ipts[cpu]; | |
360 | p = ipt->idleness; | |
361 | } | |
362 | ||
363 | return p*100.0; | |
364 | } | |
365 | ||
366 | void fio_idle_prof_cleanup(void) | |
367 | { | |
368 | if (ipc.ipts) { | |
369 | free(ipc.ipts); | |
370 | ipc.ipts = NULL; | |
371 | } | |
372 | ||
373 | if (ipc.buf) { | |
374 | free(ipc.buf); | |
375 | ipc.buf = NULL; | |
376 | } | |
377 | } | |
378 | ||
379 | int fio_idle_prof_parse_opt(const char *args) | |
380 | { | |
381 | ipc.opt = IDLE_PROF_OPT_NONE; /* default */ | |
382 | ||
383 | if (!args) { | |
384 | log_err("fio: empty idle-prof option string\n"); | |
385 | return -1; | |
386 | } | |
387 | ||
388 | #if defined(FIO_HAVE_CPU_AFFINITY) && defined(FIO_HAVE_SCHED_IDLE) | |
389 | if (strcmp("calibrate", args) == 0) { | |
390 | ipc.opt = IDLE_PROF_OPT_CALI; | |
391 | fio_idle_prof_init(); | |
392 | fio_idle_prof_start(); | |
393 | fio_idle_prof_stop(); | |
394 | show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL); | |
395 | return 1; | |
396 | } else if (strcmp("system", args) == 0) { | |
397 | ipc.opt = IDLE_PROF_OPT_SYSTEM; | |
398 | return 0; | |
399 | } else if (strcmp("percpu", args) == 0) { | |
400 | ipc.opt = IDLE_PROF_OPT_PERCPU; | |
401 | return 0; | |
402 | } else { | |
403 | log_err("fio: incorrect idle-prof option\n", args); | |
404 | return -1; | |
405 | } | |
406 | #else | |
407 | log_err("fio: idle-prof not supported on this platform\n"); | |
408 | return -1; | |
409 | #endif | |
410 | } | |
411 | ||
412 | void show_idle_prof_stats(int output, struct json_object *parent) | |
413 | { | |
414 | int i, nr_cpus = ipc.nr_cpus; | |
415 | struct json_object *tmp; | |
416 | char s[MAX_CPU_STR_LEN]; | |
417 | ||
418 | if (output == FIO_OUTPUT_NORMAL) { | |
419 | if (ipc.opt > IDLE_PROF_OPT_CALI) | |
420 | log_info("\nCPU idleness:\n"); | |
421 | else if (ipc.opt == IDLE_PROF_OPT_CALI) | |
422 | log_info("CPU idleness:\n"); | |
423 | ||
424 | if (ipc.opt >= IDLE_PROF_OPT_SYSTEM) | |
425 | log_info(" system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1)); | |
426 | ||
427 | if (ipc.opt == IDLE_PROF_OPT_PERCPU) { | |
428 | log_info(" percpu: %3.2f%%", fio_idle_prof_cpu_stat(0)); | |
429 | for (i=1; i<nr_cpus; i++) { | |
430 | log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i)); | |
431 | } | |
432 | log_info("\n"); | |
433 | } | |
434 | ||
435 | if (ipc.opt >= IDLE_PROF_OPT_CALI) { | |
436 | log_info(" unit work: mean=%3.2fus,", ipc.cali_mean); | |
437 | log_info(" stddev=%3.2f\n", ipc.cali_stddev); | |
438 | } | |
439 | ||
440 | /* dynamic mem allocations can now be freed */ | |
441 | if (ipc.opt != IDLE_PROF_OPT_NONE) | |
442 | fio_idle_prof_cleanup(); | |
443 | ||
444 | return; | |
445 | } | |
446 | ||
447 | if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) { | |
448 | if (!parent) | |
449 | return; | |
450 | ||
451 | tmp = json_create_object(); | |
452 | if (!tmp) | |
453 | return; | |
454 | ||
455 | json_object_add_value_object(parent, "cpu_idleness", tmp); | |
456 | json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1)); | |
457 | ||
458 | if (ipc.opt == IDLE_PROF_OPT_PERCPU) { | |
459 | for (i=0; i<nr_cpus; i++) { | |
460 | snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i); | |
461 | json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i)); | |
462 | } | |
463 | } | |
464 | ||
465 | json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean); | |
466 | json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev); | |
467 | ||
468 | fio_idle_prof_cleanup(); | |
469 | ||
470 | return; | |
471 | } | |
472 | } |