[fio.git] / idletime.c

#include <math.h>
#include "json.h"
#include "idletime.h"

static volatile struct idle_prof_common ipc;

/*
 * Get time to complete an unit work on a particular cpu.
 * The minimum number in CALIBRATE_RUNS runs is returned.
 */
static double calibrate_unit(unsigned char *data)
{
	unsigned long t, i, j, k;
	struct timeval tps;
	double tunit = 0.0;

	for (i = 0; i < CALIBRATE_RUNS; i++) {

		fio_gettime(&tps, NULL);
		/* scale for less variance */
		for (j = 0; j < CALIBRATE_SCALE; j++) {
			/* unit of work */
			for (k=0; k < page_size; k++) {
				data[(k + j) % page_size] = k % 256;
				/*
				 * we won't see STOP here. this is to match
				 * the same statement in the profiling loop.
				 */
				if (ipc.status == IDLE_PROF_STATUS_PROF_STOP)
					return 0.0;
			}
		}

		t = utime_since_now(&tps);
		if (!t)
			continue;

		/* get the minimum time to complete CALIBRATE_SCALE units */
		if ((i == 0) || ((double)t < tunit))
			tunit = (double)t;
	}

	return tunit / CALIBRATE_SCALE;
}

static int set_cpu_affinity(struct idle_prof_thread *ipt)
{
#if defined(FIO_HAVE_CPU_AFFINITY)
	os_cpu_mask_t cpu_mask;

	memset(&cpu_mask, 0, sizeof(cpu_mask));
	fio_cpu_set(&cpu_mask, ipt->cpu);

	if (fio_setaffinity(gettid(), cpu_mask)) {
		log_err("fio: fio_setaffinity failed\n");
		return -1;
	}

	return 0;
#else
	log_err("fio: fio_setaffinity not supported\n");
	return -1;
#endif
}

static void *idle_prof_thread_fn(void *data)
{
	int retval;
	unsigned long j, k;
	struct idle_prof_thread *ipt = data;

	/* wait for all threads are spawned */
	pthread_mutex_lock(&ipt->init_lock);

	/* exit if any other thread failed to start */
	if (ipc.status == IDLE_PROF_STATUS_ABORT)
		return NULL;

	retval = set_cpu_affinity(ipt);
	if (retval == -1) {
		ipt->state = TD_EXITED;
		pthread_mutex_unlock(&ipt->init_lock);
		return NULL;
        }

	ipt->cali_time = calibrate_unit(ipt->data);

	/* delay to set IDLE class till now for better calibration accuracy */
#if defined(CONFIG_SCHED_IDLE)
	if ((retval = fio_set_sched_idle()))
		log_err("fio: fio_set_sched_idle failed\n");
#else
	retval = -1;
	log_err("fio: fio_set_sched_idle not supported\n");
#endif
	if (retval == -1) {
		ipt->state = TD_EXITED;
		pthread_mutex_unlock(&ipt->init_lock);
		return NULL;
	}

	ipt->state = TD_INITIALIZED;

	/* signal the main thread that calibration is done */
	pthread_cond_signal(&ipt->cond);
	pthread_mutex_unlock(&ipt->init_lock);

	/* wait for other calibration to finish */
	pthread_mutex_lock(&ipt->start_lock);

	/* exit if other threads failed to initialize */
	if (ipc.status == IDLE_PROF_STATUS_ABORT)
		return NULL;

	/* exit if we are doing calibration only */
	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP)
		return NULL;

	fio_gettime(&ipt->tps, NULL);
	ipt->state = TD_RUNNING;

	j = 0;
	while (1) {
		for (k = 0; k < page_size; k++) {
			ipt->data[(k + j) % page_size] = k % 256;
			if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
				fio_gettime(&ipt->tpe, NULL);
				goto idle_prof_done;
			}
		}
		j++;
	}

idle_prof_done:

	ipt->loops = j + (double) k / page_size;
	ipt->state = TD_EXITED;
	pthread_mutex_unlock(&ipt->start_lock);

	return NULL;
}

/* calculate mean and standard deviation to complete an unit of work */
static void calibration_stats(void)
{
	int i;
	double sum = 0.0, var = 0.0;
	struct idle_prof_thread *ipt;

	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		sum += ipt->cali_time;
	}

	ipc.cali_mean = sum/ipc.nr_cpus;

	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		var += pow(ipt->cali_time-ipc.cali_mean, 2);
	}

	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
}

void fio_idle_prof_init(void)
{
	int i, ret;
	struct timeval tp;
	struct timespec ts;
	pthread_attr_t tattr;
	struct idle_prof_thread *ipt;

	ipc.nr_cpus = cpus_online();
	ipc.status = IDLE_PROF_STATUS_OK;

	if (ipc.opt == IDLE_PROF_OPT_NONE)
		return;

	if ((ret = pthread_attr_init(&tattr))) {
		log_err("fio: pthread_attr_init %s\n", strerror(ret));
		return;
	}
	if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
		log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
		return;
	}

	ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
	if (!ipc.ipts) {
		log_err("fio: malloc failed\n");
		return;
	}

	ipc.buf = malloc(ipc.nr_cpus * page_size);
	if (!ipc.buf) {
		log_err("fio: malloc failed\n");
		free(ipc.ipts);
		return;
	}

	/*
	 * profiling aborts on any single thread failure since the
	 * result won't be accurate if any cpu is not used.
	 */
	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];

		ipt->cpu = i;	
		ipt->state = TD_NOT_CREATED;
		ipt->data = (unsigned char *)(ipc.buf + page_size * i);

		if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) {
			ipc.status = IDLE_PROF_STATUS_ABORT;
			log_err("fio: pthread_mutex_init %s\n", strerror(ret));
			break;
		}

		if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) {
			ipc.status = IDLE_PROF_STATUS_ABORT;
			log_err("fio: pthread_mutex_init %s\n", strerror(ret));
			break;
		}

		if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
			ipc.status = IDLE_PROF_STATUS_ABORT;
			log_err("fio: pthread_cond_init %s\n", strerror(ret));
			break;
		}

		/* make sure all threads are spawned before they start */
		pthread_mutex_lock(&ipt->init_lock);

		/* make sure all threads finish init before profiling starts */
		pthread_mutex_lock(&ipt->start_lock);

		if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
			ipc.status = IDLE_PROF_STATUS_ABORT;
			log_err("fio: pthread_create %s\n", strerror(ret));
			break;
		} else
			ipt->state = TD_CREATED;

		if ((ret = pthread_detach(ipt->thread))) {
			/* log error and let the thread spin */
			log_err("fio: pthread_detatch %s\n", strerror(ret));
		}
	}

	/*
	 * let good threads continue so that they can exit
	 * if errors on other threads occurred previously.
	 */
	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		pthread_mutex_unlock(&ipt->init_lock);
	}
	
	if (ipc.status == IDLE_PROF_STATUS_ABORT)
		return;
	
	/* wait for calibration to finish */
	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		pthread_mutex_lock(&ipt->init_lock);
		while ((ipt->state != TD_EXITED) &&
		       (ipt->state!=TD_INITIALIZED)) {
			fio_gettime(&tp, NULL);
			ts.tv_sec = tp.tv_sec + 1;
			ts.tv_nsec = tp.tv_usec * 1000;
			pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
		}
		pthread_mutex_unlock(&ipt->init_lock);
	
		/*
		 * any thread failed to initialize would abort other threads
		 * later after fio_idle_prof_start. 
		 */	
		if (ipt->state == TD_EXITED)
			ipc.status = IDLE_PROF_STATUS_ABORT;
	}

	if (ipc.status != IDLE_PROF_STATUS_ABORT)
		calibration_stats();
	else
		ipc.cali_mean = ipc.cali_stddev = 0.0;

	if (ipc.opt == IDLE_PROF_OPT_CALI)
		ipc.status = IDLE_PROF_STATUS_CALI_STOP;
}

void fio_idle_prof_start(void)
{
	int i;
	struct idle_prof_thread *ipt;

	if (ipc.opt == IDLE_PROF_OPT_NONE)
		return;

	/* unlock regardless abort is set or not */
	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		pthread_mutex_unlock(&ipt->start_lock);
	}
}

void fio_idle_prof_stop(void)
{
	int i;
	uint64_t runt;
	struct timeval tp;
	struct timespec ts;
	struct idle_prof_thread *ipt;

	if (ipc.opt == IDLE_PROF_OPT_NONE)
		return;

	if (ipc.opt == IDLE_PROF_OPT_CALI)
		return;

	ipc.status = IDLE_PROF_STATUS_PROF_STOP;

	/* wait for all threads to exit from profiling */
	for (i = 0; i < ipc.nr_cpus; i++) {
		ipt = &ipc.ipts[i];
		pthread_mutex_lock(&ipt->start_lock);
		while ((ipt->state != TD_EXITED) &&
		       (ipt->state!=TD_NOT_CREATED)) {
			fio_gettime(&tp, NULL);
			ts.tv_sec = tp.tv_sec + 1;
			ts.tv_nsec = tp.tv_usec * 1000;
			/* timed wait in case a signal is not received */
			pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
		}
		pthread_mutex_unlock(&ipt->start_lock);

		/* calculate idleness */
		if (ipc.cali_mean != 0.0) {
			runt = utime_since(&ipt->tps, &ipt->tpe);
			ipt->idleness = ipt->loops * ipc.cali_mean / runt;
		} else
			ipt->idleness = 0.0;
	}

	/*
	 * memory allocations are freed via explicit fio_idle_prof_cleanup
	 * after profiling stats are collected by apps.  
	 */
}

/*
 * return system idle percentage when cpu is -1;
 * return one cpu idle percentage otherwise.
 */
static double fio_idle_prof_cpu_stat(int cpu)
{
	int i, nr_cpus = ipc.nr_cpus;
	struct idle_prof_thread *ipt;
	double p = 0.0;

	if (ipc.opt == IDLE_PROF_OPT_NONE)
		return 0.0;

	if ((cpu >= nr_cpus) || (cpu < -1)) {
		log_err("fio: idle profiling invalid cpu index\n");
		return 0.0;
	}

	if (cpu == -1) {
		for (i = 0; i < nr_cpus; i++) {
			ipt = &ipc.ipts[i];
			p += ipt->idleness;
		}
		p /= nr_cpus;
	} else {
		ipt = &ipc.ipts[cpu];
		p = ipt->idleness;
	}

	return p * 100.0;
}

void fio_idle_prof_cleanup(void)
{
	if (ipc.ipts) {
		free(ipc.ipts);
		ipc.ipts = NULL;
	}

	if (ipc.buf) {
		free(ipc.buf);
		ipc.buf = NULL;
	}
}

int fio_idle_prof_parse_opt(const char *args)
{
	ipc.opt = IDLE_PROF_OPT_NONE; /* default */

	if (!args) {
		log_err("fio: empty idle-prof option string\n");
		return -1;
	}	

#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
	if (strcmp("calibrate", args) == 0) {
		ipc.opt = IDLE_PROF_OPT_CALI;
		fio_idle_prof_init();
		fio_idle_prof_start();
		fio_idle_prof_stop();
		show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
		return 1;
	} else if (strcmp("system", args) == 0) {
		ipc.opt = IDLE_PROF_OPT_SYSTEM;
		return 0;
	} else if (strcmp("percpu", args) == 0) {
		ipc.opt = IDLE_PROF_OPT_PERCPU;
		return 0;
	} else {
		log_err("fio: incorrect idle-prof option: %s\n", args);
		return -1;
	}	
#else
	log_err("fio: idle-prof not supported on this platform\n");
	return -1;
#endif
}

void show_idle_prof_stats(int output, struct json_object *parent)
{
	int i, nr_cpus = ipc.nr_cpus;
	struct json_object *tmp;
	char s[MAX_CPU_STR_LEN];

	if (output == FIO_OUTPUT_NORMAL) {
		if (ipc.opt > IDLE_PROF_OPT_CALI)
			log_info("\nCPU idleness:\n");
		else if (ipc.opt == IDLE_PROF_OPT_CALI)
			log_info("CPU idleness:\n");

		if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
			log_info("  system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));

		if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
			log_info("  percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
			for (i = 1; i < nr_cpus; i++)
				log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i));
			log_info("\n");
		}

		if (ipc.opt >= IDLE_PROF_OPT_CALI) {
			log_info("  unit work: mean=%3.2fus,", ipc.cali_mean);
			log_info(" stddev=%3.2f\n", ipc.cali_stddev);
		}

		/* dynamic mem allocations can now be freed */
		if (ipc.opt != IDLE_PROF_OPT_NONE)
			fio_idle_prof_cleanup();

		return;
	}

	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) {
		if (!parent)
			return;

		tmp = json_create_object();
		if (!tmp)
			return;

		json_object_add_value_object(parent, "cpu_idleness", tmp);
		json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));

		if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
			for (i = 0; i < nr_cpus; i++) {
				snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
				json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
			}
		}

		json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
		json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
		
		fio_idle_prof_cleanup();
	}
}
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
680b5abc JA	7	/*
680b5abc JA	8	* Get time to complete an unit work on a particular cpu.
f2a2ce0e HL	9	* The minimum number in CALIBRATE_RUNS runs is returned.
	10	*/
	11	static double calibrate_unit(unsigned char *data)
	12	{
	13	unsigned long t, i, j, k;
	14	struct timeval tps;
	15	double tunit = 0.0;
	16
680b5abc	17	for (i = 0; i < CALIBRATE_RUNS; i++) {
f2a2ce0e HL	18
	19	fio_gettime(&tps, NULL);
	20	/* scale for less variance */
680b5abc	21	for (j = 0; j < CALIBRATE_SCALE; j++) {
f2a2ce0e HL	22	/* unit of work */
f2a2ce0e HL	23	for (k=0; k < page_size; k++) {
680b5abc JA	24	data[(k + j) % page_size] = k % 256;
	25	/*
	26	* we won't see STOP here. this is to match
f2a2ce0e HL	27	* the same statement in the profiling loop.
	28	*/
	29	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP)
	30	return 0.0;
	31	}
	32	}
	33
	34	t = utime_since_now(&tps);
	35	if (!t)
	36	continue;
	37
	38	/* get the minimum time to complete CALIBRATE_SCALE units */
680b5abc	39	if ((i == 0) \|\| ((double)t < tunit))
f2a2ce0e HL	40	tunit = (double)t;
	41	}
	42
680b5abc	43	return tunit / CALIBRATE_SCALE;
f2a2ce0e HL	44	}
f2a2ce0e HL	45
59358c8e JA	46	static int set_cpu_affinity(struct idle_prof_thread *ipt)
	47	{
	48	#if defined(FIO_HAVE_CPU_AFFINITY)
	49	os_cpu_mask_t cpu_mask;
	50
	51	memset(&cpu_mask, 0, sizeof(cpu_mask));
	52	fio_cpu_set(&cpu_mask, ipt->cpu);
	53
	54	if (fio_setaffinity(gettid(), cpu_mask)) {
	55	log_err("fio: fio_setaffinity failed\n");
	56	return -1;
	57	}
	58
	59	return 0;
	60	#else
	61	log_err("fio: fio_setaffinity not supported\n");
	62	return -1;
	63	#endif
	64	}
	65
f2a2ce0e HL	66	static void idle_prof_thread_fn(void data)
	67	{
	68	int retval;
	69	unsigned long j, k;
	70	struct idle_prof_thread *ipt = data;
	71
	72	/* wait for all threads are spawned */
	73	pthread_mutex_lock(&ipt->init_lock);
	74
	75	/* exit if any other thread failed to start */
	76	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	77	return NULL;
	78
59358c8e	79	retval = set_cpu_affinity(ipt);
f2a2ce0e HL	80	if (retval == -1) {
	81	ipt->state = TD_EXITED;
	82	pthread_mutex_unlock(&ipt->init_lock);
	83	return NULL;
	84	}
	85
	86	ipt->cali_time = calibrate_unit(ipt->data);
	87
	88	/* delay to set IDLE class till now for better calibration accuracy */
7e09a9f1	89	#if defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	90	if ((retval = fio_set_sched_idle()))
	91	log_err("fio: fio_set_sched_idle failed\n");
	92	#else
	93	retval = -1;
	94	log_err("fio: fio_set_sched_idle not supported\n");
	95	#endif
	96	if (retval == -1) {
	97	ipt->state = TD_EXITED;
	98	pthread_mutex_unlock(&ipt->init_lock);
	99	return NULL;
	100	}
	101
	102	ipt->state = TD_INITIALIZED;
	103
	104	/* signal the main thread that calibration is done */
	105	pthread_cond_signal(&ipt->cond);
	106	pthread_mutex_unlock(&ipt->init_lock);
	107
	108	/* wait for other calibration to finish */
	109	pthread_mutex_lock(&ipt->start_lock);
	110
	111	/* exit if other threads failed to initialize */
	112	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	113	return NULL;
	114
	115	/* exit if we are doing calibration only */
	116	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP)
	117	return NULL;
	118
	119	fio_gettime(&ipt->tps, NULL);
	120	ipt->state = TD_RUNNING;
	121
	122	j = 0;
	123	while (1) {
680b5abc JA	124	for (k = 0; k < page_size; k++) {
680b5abc JA	125	ipt->data[(k + j) % page_size] = k % 256;
f2a2ce0e HL	126	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
	127	fio_gettime(&ipt->tpe, NULL);
	128	goto idle_prof_done;
	129	}
	130	}
	131	j++;
	132	}
	133
	134	idle_prof_done:
	135
680b5abc	136	ipt->loops = j + (double) k / page_size;
f2a2ce0e HL	137	ipt->state = TD_EXITED;
	138	pthread_mutex_unlock(&ipt->start_lock);
	139
	140	return NULL;
	141	}
	142
	143	/* calculate mean and standard deviation to complete an unit of work */
	144	static void calibration_stats(void)
	145	{
	146	int i;
680b5abc	147	double sum = 0.0, var = 0.0;
f2a2ce0e HL	148	struct idle_prof_thread *ipt;
	149
	150	for (i = 0; i < ipc.nr_cpus; i++) {
	151	ipt = &ipc.ipts[i];
	152	sum += ipt->cali_time;
	153	}
	154
	155	ipc.cali_mean = sum/ipc.nr_cpus;
	156
	157	for (i = 0; i < ipc.nr_cpus; i++) {
	158	ipt = &ipc.ipts[i];
	159	var += pow(ipt->cali_time-ipc.cali_mean, 2);
	160	}
	161
	162	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
	163	}
	164
	165	void fio_idle_prof_init(void)
	166	{
	167	int i, ret;
	168	struct timeval tp;
	169	struct timespec ts;
680b5abc	170	pthread_attr_t tattr;
f2a2ce0e HL	171	struct idle_prof_thread *ipt;
	172
	173	ipc.nr_cpus = cpus_online();
	174	ipc.status = IDLE_PROF_STATUS_OK;
	175
	176	if (ipc.opt == IDLE_PROF_OPT_NONE)
	177	return;
	178
	179	if ((ret = pthread_attr_init(&tattr))) {
	180	log_err("fio: pthread_attr_init %s\n", strerror(ret));
	181	return;
	182	}
	183	if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
	184	log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
	185	return;
	186	}
	187
	188	ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
	189	if (!ipc.ipts) {
	190	log_err("fio: malloc failed\n");
	191	return;
	192	}
	193
	194	ipc.buf = malloc(ipc.nr_cpus * page_size);
	195	if (!ipc.buf) {
	196	log_err("fio: malloc failed\n");
	197	free(ipc.ipts);
	198	return;
	199	}
	200
680b5abc JA	201	/*
680b5abc JA	202	* profiling aborts on any single thread failure since the
f2a2ce0e HL	203	* result won't be accurate if any cpu is not used.
	204	*/
	205	for (i = 0; i < ipc.nr_cpus; i++) {
	206	ipt = &ipc.ipts[i];
	207
	208	ipt->cpu = i;
	209	ipt->state = TD_NOT_CREATED;
	210	ipt->data = (unsigned char )(ipc.buf + page_size i);
	211
	212	if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) {
	213	ipc.status = IDLE_PROF_STATUS_ABORT;
	214	log_err("fio: pthread_mutex_init %s\n", strerror(ret));
	215	break;
	216	}
	217
	218	if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) {
	219	ipc.status = IDLE_PROF_STATUS_ABORT;
	220	log_err("fio: pthread_mutex_init %s\n", strerror(ret));
	221	break;
	222	}
	223
	224	if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
	225	ipc.status = IDLE_PROF_STATUS_ABORT;
	226	log_err("fio: pthread_cond_init %s\n", strerror(ret));
	227	break;
	228	}
	229
	230	/* make sure all threads are spawned before they start */
	231	pthread_mutex_lock(&ipt->init_lock);
	232
	233	/* make sure all threads finish init before profiling starts */
	234	pthread_mutex_lock(&ipt->start_lock);
	235
	236	if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
	237	ipc.status = IDLE_PROF_STATUS_ABORT;
	238	log_err("fio: pthread_create %s\n", strerror(ret));
	239	break;
680b5abc	240	} else
f2a2ce0e	241	ipt->state = TD_CREATED;
f2a2ce0e HL	242
	243	if ((ret = pthread_detach(ipt->thread))) {
	244	/* log error and let the thread spin */
	245	log_err("fio: pthread_detatch %s\n", strerror(ret));
	246	}
	247	}
	248
680b5abc JA	249	/*
	250	* let good threads continue so that they can exit
	251	* if errors on other threads occurred previously.
f2a2ce0e HL	252	*/
	253	for (i = 0; i < ipc.nr_cpus; i++) {
	254	ipt = &ipc.ipts[i];
	255	pthread_mutex_unlock(&ipt->init_lock);
	256	}
	257
	258	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	259	return;
	260
	261	/* wait for calibration to finish */
	262	for (i = 0; i < ipc.nr_cpus; i++) {
	263	ipt = &ipc.ipts[i];
	264	pthread_mutex_lock(&ipt->init_lock);
680b5abc JA	265	while ((ipt->state != TD_EXITED) &&
680b5abc JA	266	(ipt->state!=TD_INITIALIZED)) {
f2a2ce0e HL	267	fio_gettime(&tp, NULL);
	268	ts.tv_sec = tp.tv_sec + 1;
	269	ts.tv_nsec = tp.tv_usec * 1000;
	270	pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
	271	}
	272	pthread_mutex_unlock(&ipt->init_lock);
	273
680b5abc JA	274	/*
680b5abc JA	275	* any thread failed to initialize would abort other threads
f2a2ce0e HL	276	* later after fio_idle_prof_start.
	277	*/
	278	if (ipt->state == TD_EXITED)
	279	ipc.status = IDLE_PROF_STATUS_ABORT;
	280	}
	281
	282	if (ipc.status != IDLE_PROF_STATUS_ABORT)
	283	calibration_stats();
680b5abc	284	else
f2a2ce0e HL	285	ipc.cali_mean = ipc.cali_stddev = 0.0;
	286
	287	if (ipc.opt == IDLE_PROF_OPT_CALI)
	288	ipc.status = IDLE_PROF_STATUS_CALI_STOP;
	289	}
	290
	291	void fio_idle_prof_start(void)
	292	{
	293	int i;
	294	struct idle_prof_thread *ipt;
	295
	296	if (ipc.opt == IDLE_PROF_OPT_NONE)
	297	return;
	298
	299	/* unlock regardless abort is set or not */
	300	for (i = 0; i < ipc.nr_cpus; i++) {
	301	ipt = &ipc.ipts[i];
	302	pthread_mutex_unlock(&ipt->start_lock);
	303	}
	304	}
	305
	306	void fio_idle_prof_stop(void)
	307	{
	308	int i;
	309	uint64_t runt;
	310	struct timeval tp;
	311	struct timespec ts;
	312	struct idle_prof_thread *ipt;
	313
	314	if (ipc.opt == IDLE_PROF_OPT_NONE)
	315	return;
	316
	317	if (ipc.opt == IDLE_PROF_OPT_CALI)
	318	return;
	319
	320	ipc.status = IDLE_PROF_STATUS_PROF_STOP;
	321
	322	/* wait for all threads to exit from profiling */
	323	for (i = 0; i < ipc.nr_cpus; i++) {
	324	ipt = &ipc.ipts[i];
	325	pthread_mutex_lock(&ipt->start_lock);
680b5abc JA	326	while ((ipt->state != TD_EXITED) &&
680b5abc JA	327	(ipt->state!=TD_NOT_CREATED)) {
f2a2ce0e HL	328	fio_gettime(&tp, NULL);
	329	ts.tv_sec = tp.tv_sec + 1;
	330	ts.tv_nsec = tp.tv_usec * 1000;
	331	/* timed wait in case a signal is not received */
	332	pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
	333	}
	334	pthread_mutex_unlock(&ipt->start_lock);
	335
	336	/* calculate idleness */
	337	if (ipc.cali_mean != 0.0) {
	338	runt = utime_since(&ipt->tps, &ipt->tpe);
	339	ipt->idleness = ipt->loops * ipc.cali_mean / runt;
680b5abc	340	} else
f2a2ce0e HL	341	ipt->idleness = 0.0;
	342	}
	343
680b5abc JA	344	/*
680b5abc JA	345	* memory allocations are freed via explicit fio_idle_prof_cleanup
f2a2ce0e HL	346	* after profiling stats are collected by apps.
f2a2ce0e HL	347	*/
f2a2ce0e HL	348	}
f2a2ce0e HL	349
680b5abc JA	350	/*
680b5abc JA	351	* return system idle percentage when cpu is -1;
f2a2ce0e HL	352	* return one cpu idle percentage otherwise.
	353	*/
	354	static double fio_idle_prof_cpu_stat(int cpu)
	355	{
	356	int i, nr_cpus = ipc.nr_cpus;
	357	struct idle_prof_thread *ipt;
	358	double p = 0.0;
	359
	360	if (ipc.opt == IDLE_PROF_OPT_NONE)
	361	return 0.0;
	362
	363	if ((cpu >= nr_cpus) \|\| (cpu < -1)) {
	364	log_err("fio: idle profiling invalid cpu index\n");
	365	return 0.0;
	366	}
	367
	368	if (cpu == -1) {
	369	for (i = 0; i < nr_cpus; i++) {
	370	ipt = &ipc.ipts[i];
	371	p += ipt->idleness;
	372	}
	373	p /= nr_cpus;
	374	} else {
	375	ipt = &ipc.ipts[cpu];
	376	p = ipt->idleness;
	377	}
	378
680b5abc	379	return p * 100.0;
f2a2ce0e HL	380	}
	381
	382	void fio_idle_prof_cleanup(void)
	383	{
	384	if (ipc.ipts) {
	385	free(ipc.ipts);
	386	ipc.ipts = NULL;
	387	}
	388
	389	if (ipc.buf) {
	390	free(ipc.buf);
	391	ipc.buf = NULL;
	392	}
	393	}
	394
	395	int fio_idle_prof_parse_opt(const char *args)
	396	{
	397	ipc.opt = IDLE_PROF_OPT_NONE; /* default */
	398
	399	if (!args) {
	400	log_err("fio: empty idle-prof option string\n");
	401	return -1;
	402	}
	403
7e09a9f1	404	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	405	if (strcmp("calibrate", args) == 0) {
	406	ipc.opt = IDLE_PROF_OPT_CALI;
	407	fio_idle_prof_init();
	408	fio_idle_prof_start();
	409	fio_idle_prof_stop();
	410	show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
	411	return 1;
	412	} else if (strcmp("system", args) == 0) {
	413	ipc.opt = IDLE_PROF_OPT_SYSTEM;
	414	return 0;
	415	} else if (strcmp("percpu", args) == 0) {
	416	ipc.opt = IDLE_PROF_OPT_PERCPU;
	417	return 0;
	418	} else {
4e0a8fa2	419	log_err("fio: incorrect idle-prof option: %s\n", args);
f2a2ce0e HL	420	return -1;
	421	}
	422	#else
	423	log_err("fio: idle-prof not supported on this platform\n");
	424	return -1;
	425	#endif
	426	}
	427
	428	void show_idle_prof_stats(int output, struct json_object *parent)
	429	{
	430	int i, nr_cpus = ipc.nr_cpus;
	431	struct json_object *tmp;
	432	char s[MAX_CPU_STR_LEN];
680b5abc	433
f2a2ce0e HL	434	if (output == FIO_OUTPUT_NORMAL) {
	435	if (ipc.opt > IDLE_PROF_OPT_CALI)
	436	log_info("\nCPU idleness:\n");
	437	else if (ipc.opt == IDLE_PROF_OPT_CALI)
	438	log_info("CPU idleness:\n");
	439
	440	if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
	441	log_info(" system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
	442
	443	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
	444	log_info(" percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
680b5abc	445	for (i = 1; i < nr_cpus; i++)
f2a2ce0e	446	log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i));
f2a2ce0e HL	447	log_info("\n");
	448	}
	449
	450	if (ipc.opt >= IDLE_PROF_OPT_CALI) {
	451	log_info(" unit work: mean=%3.2fus,", ipc.cali_mean);
	452	log_info(" stddev=%3.2f\n", ipc.cali_stddev);
	453	}
	454
	455	/* dynamic mem allocations can now be freed */
	456	if (ipc.opt != IDLE_PROF_OPT_NONE)
	457	fio_idle_prof_cleanup();
	458
	459	return;
	460	}
680b5abc	461
f2a2ce0e HL	462	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) {
	463	if (!parent)
	464	return;
	465
	466	tmp = json_create_object();
	467	if (!tmp)
	468	return;
	469
	470	json_object_add_value_object(parent, "cpu_idleness", tmp);
	471	json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
	472
	473	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
680b5abc	474	for (i = 0; i < nr_cpus; i++) {
f2a2ce0e HL	475	snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
	476	json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
	477	}
	478	}
	479
	480	json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
	481	json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
	482
	483	fio_idle_prof_cleanup();
f2a2ce0e HL	484	}
f2a2ce0e HL	485	}