5 static volatile struct idle_prof_common ipc;
8 * Get time to complete an unit work on a particular cpu.
9 * The minimum number in CALIBRATE_RUNS runs is returned.
11 static double calibrate_unit(unsigned char *data)
13 unsigned long t, i, j, k;
17 for (i = 0; i < CALIBRATE_RUNS; i++) {
19 fio_gettime(&tps, NULL);
20 /* scale for less variance */
21 for (j = 0; j < CALIBRATE_SCALE; j++) {
23 for (k=0; k < page_size; k++) {
24 data[(k + j) % page_size] = k % 256;
26 * we won't see STOP here. this is to match
27 * the same statement in the profiling loop.
29 if (ipc.status == IDLE_PROF_STATUS_PROF_STOP)
34 t = utime_since_now(&tps);
38 /* get the minimum time to complete CALIBRATE_SCALE units */
39 if ((i == 0) || ((double)t < tunit))
43 return tunit / CALIBRATE_SCALE;
46 static void free_cpu_affinity(struct idle_prof_thread *ipt)
48 #if defined(FIO_HAVE_CPU_AFFINITY)
49 fio_cpuset_exit(&ipt->cpu_mask);
53 static int set_cpu_affinity(struct idle_prof_thread *ipt)
55 #if defined(FIO_HAVE_CPU_AFFINITY)
56 if (fio_cpuset_init(&ipt->cpu_mask)) {
57 log_err("fio: cpuset init failed\n");
61 fio_cpu_set(&ipt->cpu_mask, ipt->cpu);
63 if (fio_setaffinity(gettid(), ipt->cpu_mask)) {
64 log_err("fio: fio_setaffinity failed\n");
65 fio_cpuset_exit(&ipt->cpu_mask);
71 log_err("fio: fio_setaffinity not supported\n");
76 static void *idle_prof_thread_fn(void *data)
80 struct idle_prof_thread *ipt = data;
82 /* wait for all threads are spawned */
83 pthread_mutex_lock(&ipt->init_lock);
85 /* exit if any other thread failed to start */
86 if (ipc.status == IDLE_PROF_STATUS_ABORT) {
87 pthread_mutex_unlock(&ipt->init_lock);
91 retval = set_cpu_affinity(ipt);
93 ipt->state = TD_EXITED;
94 pthread_mutex_unlock(&ipt->init_lock);
98 ipt->cali_time = calibrate_unit(ipt->data);
100 /* delay to set IDLE class till now for better calibration accuracy */
101 #if defined(CONFIG_SCHED_IDLE)
102 if ((retval = fio_set_sched_idle()))
103 log_err("fio: fio_set_sched_idle failed\n");
106 log_err("fio: fio_set_sched_idle not supported\n");
109 ipt->state = TD_EXITED;
110 pthread_mutex_unlock(&ipt->init_lock);
114 ipt->state = TD_INITIALIZED;
116 /* signal the main thread that calibration is done */
117 pthread_cond_signal(&ipt->cond);
118 pthread_mutex_unlock(&ipt->init_lock);
120 /* wait for other calibration to finish */
121 pthread_mutex_lock(&ipt->start_lock);
123 /* exit if other threads failed to initialize */
124 if (ipc.status == IDLE_PROF_STATUS_ABORT) {
125 pthread_mutex_unlock(&ipt->start_lock);
129 /* exit if we are doing calibration only */
130 if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
131 pthread_mutex_unlock(&ipt->start_lock);
135 fio_gettime(&ipt->tps, NULL);
136 ipt->state = TD_RUNNING;
140 for (k = 0; k < page_size; k++) {
141 ipt->data[(k + j) % page_size] = k % 256;
142 if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
143 fio_gettime(&ipt->tpe, NULL);
152 ipt->loops = j + (double) k / page_size;
153 ipt->state = TD_EXITED;
154 pthread_mutex_unlock(&ipt->start_lock);
157 free_cpu_affinity(ipt);
161 /* calculate mean and standard deviation to complete an unit of work */
162 static void calibration_stats(void)
165 double sum = 0.0, var = 0.0;
166 struct idle_prof_thread *ipt;
168 for (i = 0; i < ipc.nr_cpus; i++) {
170 sum += ipt->cali_time;
173 ipc.cali_mean = sum/ipc.nr_cpus;
175 for (i = 0; i < ipc.nr_cpus; i++) {
177 var += pow(ipt->cali_time-ipc.cali_mean, 2);
180 ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
183 void fio_idle_prof_init(void)
187 pthread_attr_t tattr;
188 struct idle_prof_thread *ipt;
190 ipc.nr_cpus = cpus_online();
191 ipc.status = IDLE_PROF_STATUS_OK;
193 if (ipc.opt == IDLE_PROF_OPT_NONE)
196 if ((ret = pthread_attr_init(&tattr))) {
197 log_err("fio: pthread_attr_init %s\n", strerror(ret));
200 if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
201 log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
205 ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
207 log_err("fio: malloc failed\n");
211 ipc.buf = malloc(ipc.nr_cpus * page_size);
213 log_err("fio: malloc failed\n");
219 * profiling aborts on any single thread failure since the
220 * result won't be accurate if any cpu is not used.
222 for (i = 0; i < ipc.nr_cpus; i++) {
226 ipt->state = TD_NOT_CREATED;
227 ipt->data = (unsigned char *)(ipc.buf + page_size * i);
229 if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) {
230 ipc.status = IDLE_PROF_STATUS_ABORT;
231 log_err("fio: pthread_mutex_init %s\n", strerror(ret));
235 if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) {
236 ipc.status = IDLE_PROF_STATUS_ABORT;
237 log_err("fio: pthread_mutex_init %s\n", strerror(ret));
241 if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
242 ipc.status = IDLE_PROF_STATUS_ABORT;
243 log_err("fio: pthread_cond_init %s\n", strerror(ret));
247 /* make sure all threads are spawned before they start */
248 pthread_mutex_lock(&ipt->init_lock);
250 /* make sure all threads finish init before profiling starts */
251 pthread_mutex_lock(&ipt->start_lock);
253 if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
254 ipc.status = IDLE_PROF_STATUS_ABORT;
255 log_err("fio: pthread_create %s\n", strerror(ret));
258 ipt->state = TD_CREATED;
260 if ((ret = pthread_detach(ipt->thread))) {
261 /* log error and let the thread spin */
262 log_err("fio: pthread_detach %s\n", strerror(ret));
267 * let good threads continue so that they can exit
268 * if errors on other threads occurred previously.
270 for (i = 0; i < ipc.nr_cpus; i++) {
272 pthread_mutex_unlock(&ipt->init_lock);
275 if (ipc.status == IDLE_PROF_STATUS_ABORT)
278 /* wait for calibration to finish */
279 for (i = 0; i < ipc.nr_cpus; i++) {
281 pthread_mutex_lock(&ipt->init_lock);
282 while ((ipt->state != TD_EXITED) &&
283 (ipt->state!=TD_INITIALIZED)) {
284 fio_gettime(&ts, NULL);
286 pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
288 pthread_mutex_unlock(&ipt->init_lock);
291 * any thread failed to initialize would abort other threads
292 * later after fio_idle_prof_start.
294 if (ipt->state == TD_EXITED)
295 ipc.status = IDLE_PROF_STATUS_ABORT;
298 if (ipc.status != IDLE_PROF_STATUS_ABORT)
301 ipc.cali_mean = ipc.cali_stddev = 0.0;
303 if (ipc.opt == IDLE_PROF_OPT_CALI)
304 ipc.status = IDLE_PROF_STATUS_CALI_STOP;
307 void fio_idle_prof_start(void)
310 struct idle_prof_thread *ipt;
312 if (ipc.opt == IDLE_PROF_OPT_NONE)
315 /* unlock regardless abort is set or not */
316 for (i = 0; i < ipc.nr_cpus; i++) {
318 pthread_mutex_unlock(&ipt->start_lock);
322 void fio_idle_prof_stop(void)
327 struct idle_prof_thread *ipt;
329 if (ipc.opt == IDLE_PROF_OPT_NONE)
332 if (ipc.opt == IDLE_PROF_OPT_CALI)
335 ipc.status = IDLE_PROF_STATUS_PROF_STOP;
337 /* wait for all threads to exit from profiling */
338 for (i = 0; i < ipc.nr_cpus; i++) {
340 pthread_mutex_lock(&ipt->start_lock);
341 while ((ipt->state != TD_EXITED) &&
342 (ipt->state!=TD_NOT_CREATED)) {
343 fio_gettime(&ts, NULL);
345 /* timed wait in case a signal is not received */
346 pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
348 pthread_mutex_unlock(&ipt->start_lock);
350 /* calculate idleness */
351 if (ipc.cali_mean != 0.0) {
352 runt = utime_since(&ipt->tps, &ipt->tpe);
354 ipt->idleness = ipt->loops * ipc.cali_mean / runt;
362 * memory allocations are freed via explicit fio_idle_prof_cleanup
363 * after profiling stats are collected by apps.
368 * return system idle percentage when cpu is -1;
369 * return one cpu idle percentage otherwise.
371 static double fio_idle_prof_cpu_stat(int cpu)
373 int i, nr_cpus = ipc.nr_cpus;
374 struct idle_prof_thread *ipt;
377 if (ipc.opt == IDLE_PROF_OPT_NONE)
380 if ((cpu >= nr_cpus) || (cpu < -1)) {
381 log_err("fio: idle profiling invalid cpu index\n");
386 for (i = 0; i < nr_cpus; i++) {
392 ipt = &ipc.ipts[cpu];
399 static void fio_idle_prof_cleanup(void)
412 int fio_idle_prof_parse_opt(const char *args)
414 ipc.opt = IDLE_PROF_OPT_NONE; /* default */
417 log_err("fio: empty idle-prof option string\n");
421 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
422 if (strcmp("calibrate", args) == 0) {
423 ipc.opt = IDLE_PROF_OPT_CALI;
424 fio_idle_prof_init();
425 fio_idle_prof_start();
426 fio_idle_prof_stop();
427 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, NULL);
429 } else if (strcmp("system", args) == 0) {
430 ipc.opt = IDLE_PROF_OPT_SYSTEM;
432 } else if (strcmp("percpu", args) == 0) {
433 ipc.opt = IDLE_PROF_OPT_PERCPU;
436 log_err("fio: incorrect idle-prof option: %s\n", args);
440 log_err("fio: idle-prof not supported on this platform\n");
445 void show_idle_prof_stats(int output, struct json_object *parent,
446 struct buf_output *out)
448 int i, nr_cpus = ipc.nr_cpus;
449 struct json_object *tmp;
450 char s[MAX_CPU_STR_LEN];
452 if (output == FIO_OUTPUT_NORMAL) {
453 if (ipc.opt > IDLE_PROF_OPT_CALI)
454 log_buf(out, "\nCPU idleness:\n");
455 else if (ipc.opt == IDLE_PROF_OPT_CALI)
456 log_buf(out, "CPU idleness:\n");
458 if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
459 log_buf(out, " system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
461 if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
462 log_buf(out, " percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
463 for (i = 1; i < nr_cpus; i++)
464 log_buf(out, ", %3.2f%%", fio_idle_prof_cpu_stat(i));
468 if (ipc.opt >= IDLE_PROF_OPT_CALI) {
469 log_buf(out, " unit work: mean=%3.2fus,", ipc.cali_mean);
470 log_buf(out, " stddev=%3.2f\n", ipc.cali_stddev);
473 /* dynamic mem allocations can now be freed */
474 if (ipc.opt != IDLE_PROF_OPT_NONE)
475 fio_idle_prof_cleanup();
480 if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output & FIO_OUTPUT_JSON)) {
484 tmp = json_create_object();
488 json_object_add_value_object(parent, "cpu_idleness", tmp);
489 json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
491 if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
492 for (i = 0; i < nr_cpus; i++) {
493 snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
494 json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
498 json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
499 json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
501 fio_idle_prof_cleanup();