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 int set_cpu_affinity(struct idle_prof_thread *ipt)
48 #if defined(FIO_HAVE_CPU_AFFINITY)
49 os_cpu_mask_t cpu_mask;
51 memset(&cpu_mask, 0, sizeof(cpu_mask));
52 fio_cpu_set(&cpu_mask, ipt->cpu);
54 if (fio_setaffinity(gettid(), cpu_mask)) {
55 log_err("fio: fio_setaffinity failed\n");
61 log_err("fio: fio_setaffinity not supported\n");
66 static void *idle_prof_thread_fn(void *data)
70 struct idle_prof_thread *ipt = data;
72 /* wait for all threads are spawned */
73 pthread_mutex_lock(&ipt->init_lock);
75 /* exit if any other thread failed to start */
76 if (ipc.status == IDLE_PROF_STATUS_ABORT)
79 retval = set_cpu_affinity(ipt);
81 ipt->state = TD_EXITED;
82 pthread_mutex_unlock(&ipt->init_lock);
86 ipt->cali_time = calibrate_unit(ipt->data);
88 /* delay to set IDLE class till now for better calibration accuracy */
89 #if defined(CONFIG_SCHED_IDLE)
90 if ((retval = fio_set_sched_idle()))
91 log_err("fio: fio_set_sched_idle failed\n");
94 log_err("fio: fio_set_sched_idle not supported\n");
97 ipt->state = TD_EXITED;
98 pthread_mutex_unlock(&ipt->init_lock);
102 ipt->state = TD_INITIALIZED;
104 /* signal the main thread that calibration is done */
105 pthread_cond_signal(&ipt->cond);
106 pthread_mutex_unlock(&ipt->init_lock);
108 /* wait for other calibration to finish */
109 pthread_mutex_lock(&ipt->start_lock);
111 /* exit if other threads failed to initialize */
112 if (ipc.status == IDLE_PROF_STATUS_ABORT) {
113 pthread_mutex_unlock(&ipt->start_lock);
117 /* exit if we are doing calibration only */
118 if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
119 pthread_mutex_unlock(&ipt->start_lock);
123 fio_gettime(&ipt->tps, NULL);
124 ipt->state = TD_RUNNING;
128 for (k = 0; k < page_size; k++) {
129 ipt->data[(k + j) % page_size] = k % 256;
130 if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
131 fio_gettime(&ipt->tpe, NULL);
140 ipt->loops = j + (double) k / page_size;
141 ipt->state = TD_EXITED;
142 pthread_mutex_unlock(&ipt->start_lock);
147 /* calculate mean and standard deviation to complete an unit of work */
148 static void calibration_stats(void)
151 double sum = 0.0, var = 0.0;
152 struct idle_prof_thread *ipt;
154 for (i = 0; i < ipc.nr_cpus; i++) {
156 sum += ipt->cali_time;
159 ipc.cali_mean = sum/ipc.nr_cpus;
161 for (i = 0; i < ipc.nr_cpus; i++) {
163 var += pow(ipt->cali_time-ipc.cali_mean, 2);
166 ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
169 void fio_idle_prof_init(void)
174 pthread_attr_t tattr;
175 struct idle_prof_thread *ipt;
177 ipc.nr_cpus = cpus_online();
178 ipc.status = IDLE_PROF_STATUS_OK;
180 if (ipc.opt == IDLE_PROF_OPT_NONE)
183 if ((ret = pthread_attr_init(&tattr))) {
184 log_err("fio: pthread_attr_init %s\n", strerror(ret));
187 if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
188 log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
192 ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
194 log_err("fio: malloc failed\n");
198 ipc.buf = malloc(ipc.nr_cpus * page_size);
200 log_err("fio: malloc failed\n");
206 * profiling aborts on any single thread failure since the
207 * result won't be accurate if any cpu is not used.
209 for (i = 0; i < ipc.nr_cpus; i++) {
213 ipt->state = TD_NOT_CREATED;
214 ipt->data = (unsigned char *)(ipc.buf + page_size * i);
216 if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) {
217 ipc.status = IDLE_PROF_STATUS_ABORT;
218 log_err("fio: pthread_mutex_init %s\n", strerror(ret));
222 if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) {
223 ipc.status = IDLE_PROF_STATUS_ABORT;
224 log_err("fio: pthread_mutex_init %s\n", strerror(ret));
228 if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
229 ipc.status = IDLE_PROF_STATUS_ABORT;
230 log_err("fio: pthread_cond_init %s\n", strerror(ret));
234 /* make sure all threads are spawned before they start */
235 pthread_mutex_lock(&ipt->init_lock);
237 /* make sure all threads finish init before profiling starts */
238 pthread_mutex_lock(&ipt->start_lock);
240 if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
241 ipc.status = IDLE_PROF_STATUS_ABORT;
242 log_err("fio: pthread_create %s\n", strerror(ret));
245 ipt->state = TD_CREATED;
247 if ((ret = pthread_detach(ipt->thread))) {
248 /* log error and let the thread spin */
249 log_err("fio: pthread_detatch %s\n", strerror(ret));
254 * let good threads continue so that they can exit
255 * if errors on other threads occurred previously.
257 for (i = 0; i < ipc.nr_cpus; i++) {
259 pthread_mutex_unlock(&ipt->init_lock);
262 if (ipc.status == IDLE_PROF_STATUS_ABORT)
265 /* wait for calibration to finish */
266 for (i = 0; i < ipc.nr_cpus; i++) {
268 pthread_mutex_lock(&ipt->init_lock);
269 while ((ipt->state != TD_EXITED) &&
270 (ipt->state!=TD_INITIALIZED)) {
271 fio_gettime(&tp, NULL);
272 ts.tv_sec = tp.tv_sec + 1;
273 ts.tv_nsec = tp.tv_usec * 1000;
274 pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
276 pthread_mutex_unlock(&ipt->init_lock);
279 * any thread failed to initialize would abort other threads
280 * later after fio_idle_prof_start.
282 if (ipt->state == TD_EXITED)
283 ipc.status = IDLE_PROF_STATUS_ABORT;
286 if (ipc.status != IDLE_PROF_STATUS_ABORT)
289 ipc.cali_mean = ipc.cali_stddev = 0.0;
291 if (ipc.opt == IDLE_PROF_OPT_CALI)
292 ipc.status = IDLE_PROF_STATUS_CALI_STOP;
295 void fio_idle_prof_start(void)
298 struct idle_prof_thread *ipt;
300 if (ipc.opt == IDLE_PROF_OPT_NONE)
303 /* unlock regardless abort is set or not */
304 for (i = 0; i < ipc.nr_cpus; i++) {
306 pthread_mutex_unlock(&ipt->start_lock);
310 void fio_idle_prof_stop(void)
316 struct idle_prof_thread *ipt;
318 if (ipc.opt == IDLE_PROF_OPT_NONE)
321 if (ipc.opt == IDLE_PROF_OPT_CALI)
324 ipc.status = IDLE_PROF_STATUS_PROF_STOP;
326 /* wait for all threads to exit from profiling */
327 for (i = 0; i < ipc.nr_cpus; i++) {
329 pthread_mutex_lock(&ipt->start_lock);
330 while ((ipt->state != TD_EXITED) &&
331 (ipt->state!=TD_NOT_CREATED)) {
332 fio_gettime(&tp, NULL);
333 ts.tv_sec = tp.tv_sec + 1;
334 ts.tv_nsec = tp.tv_usec * 1000;
335 /* timed wait in case a signal is not received */
336 pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
338 pthread_mutex_unlock(&ipt->start_lock);
340 /* calculate idleness */
341 if (ipc.cali_mean != 0.0) {
342 runt = utime_since(&ipt->tps, &ipt->tpe);
344 ipt->idleness = ipt->loops * ipc.cali_mean / runt;
352 * memory allocations are freed via explicit fio_idle_prof_cleanup
353 * after profiling stats are collected by apps.
358 * return system idle percentage when cpu is -1;
359 * return one cpu idle percentage otherwise.
361 static double fio_idle_prof_cpu_stat(int cpu)
363 int i, nr_cpus = ipc.nr_cpus;
364 struct idle_prof_thread *ipt;
367 if (ipc.opt == IDLE_PROF_OPT_NONE)
370 if ((cpu >= nr_cpus) || (cpu < -1)) {
371 log_err("fio: idle profiling invalid cpu index\n");
376 for (i = 0; i < nr_cpus; i++) {
382 ipt = &ipc.ipts[cpu];
389 static void fio_idle_prof_cleanup(void)
402 int fio_idle_prof_parse_opt(const char *args)
404 ipc.opt = IDLE_PROF_OPT_NONE; /* default */
407 log_err("fio: empty idle-prof option string\n");
411 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
412 if (strcmp("calibrate", args) == 0) {
413 ipc.opt = IDLE_PROF_OPT_CALI;
414 fio_idle_prof_init();
415 fio_idle_prof_start();
416 fio_idle_prof_stop();
417 show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
419 } else if (strcmp("system", args) == 0) {
420 ipc.opt = IDLE_PROF_OPT_SYSTEM;
422 } else if (strcmp("percpu", args) == 0) {
423 ipc.opt = IDLE_PROF_OPT_PERCPU;
426 log_err("fio: incorrect idle-prof option: %s\n", args);
430 log_err("fio: idle-prof not supported on this platform\n");
435 void show_idle_prof_stats(int output, struct json_object *parent)
437 int i, nr_cpus = ipc.nr_cpus;
438 struct json_object *tmp;
439 char s[MAX_CPU_STR_LEN];
441 if (output == FIO_OUTPUT_NORMAL) {
442 if (ipc.opt > IDLE_PROF_OPT_CALI)
443 log_info("\nCPU idleness:\n");
444 else if (ipc.opt == IDLE_PROF_OPT_CALI)
445 log_info("CPU idleness:\n");
447 if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
448 log_info(" system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
450 if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
451 log_info(" percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
452 for (i = 1; i < nr_cpus; i++)
453 log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i));
457 if (ipc.opt >= IDLE_PROF_OPT_CALI) {
458 log_info(" unit work: mean=%3.2fus,", ipc.cali_mean);
459 log_info(" stddev=%3.2f\n", ipc.cali_stddev);
462 /* dynamic mem allocations can now be freed */
463 if (ipc.opt != IDLE_PROF_OPT_NONE)
464 fio_idle_prof_cleanup();
469 if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) {
473 tmp = json_create_object();
477 json_object_add_value_object(parent, "cpu_idleness", tmp);
478 json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
480 if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
481 for (i = 0; i < nr_cpus; i++) {
482 snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
483 json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
487 json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
488 json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
490 fio_idle_prof_cleanup();