[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) {
		pthread_mutex_unlock(&ipt->init_lock);
		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) {
		pthread_mutex_unlock(&ipt->start_lock);
		return NULL;
	}

	/* exit if we are doing calibration only */
	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
		pthread_mutex_unlock(&ipt->start_lock);
		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);
			if (runt)
				ipt->idleness = ipt->loops * ipc.cali_mean / runt;
			else
				ipt->idleness = 0.0;
		} 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;
}

static 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 */
d5b351d4 JA	76	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
d5b351d4 JA	77	pthread_mutex_unlock(&ipt->init_lock);
f2a2ce0e	78	return NULL;
d5b351d4	79	}
f2a2ce0e	80
59358c8e	81	retval = set_cpu_affinity(ipt);
f2a2ce0e HL	82	if (retval == -1) {
	83	ipt->state = TD_EXITED;
	84	pthread_mutex_unlock(&ipt->init_lock);
	85	return NULL;
	86	}
	87
	88	ipt->cali_time = calibrate_unit(ipt->data);
	89
	90	/* delay to set IDLE class till now for better calibration accuracy */
7e09a9f1	91	#if defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	92	if ((retval = fio_set_sched_idle()))
	93	log_err("fio: fio_set_sched_idle failed\n");
	94	#else
	95	retval = -1;
	96	log_err("fio: fio_set_sched_idle not supported\n");
	97	#endif
	98	if (retval == -1) {
	99	ipt->state = TD_EXITED;
	100	pthread_mutex_unlock(&ipt->init_lock);
	101	return NULL;
	102	}
	103
	104	ipt->state = TD_INITIALIZED;
	105
	106	/* signal the main thread that calibration is done */
	107	pthread_cond_signal(&ipt->cond);
	108	pthread_mutex_unlock(&ipt->init_lock);
	109
	110	/* wait for other calibration to finish */
	111	pthread_mutex_lock(&ipt->start_lock);
	112
	113	/* exit if other threads failed to initialize */
735ed278 JA	114	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
735ed278 JA	115	pthread_mutex_unlock(&ipt->start_lock);
f2a2ce0e	116	return NULL;
735ed278	117	}
f2a2ce0e HL	118
f2a2ce0e HL	119	/* exit if we are doing calibration only */
735ed278 JA	120	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
735ed278 JA	121	pthread_mutex_unlock(&ipt->start_lock);
f2a2ce0e	122	return NULL;
735ed278	123	}
f2a2ce0e HL	124
	125	fio_gettime(&ipt->tps, NULL);
	126	ipt->state = TD_RUNNING;
	127
	128	j = 0;
	129	while (1) {
680b5abc JA	130	for (k = 0; k < page_size; k++) {
680b5abc JA	131	ipt->data[(k + j) % page_size] = k % 256;
f2a2ce0e HL	132	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
	133	fio_gettime(&ipt->tpe, NULL);
	134	goto idle_prof_done;
	135	}
	136	}
	137	j++;
	138	}
	139
	140	idle_prof_done:
	141
680b5abc	142	ipt->loops = j + (double) k / page_size;
f2a2ce0e HL	143	ipt->state = TD_EXITED;
	144	pthread_mutex_unlock(&ipt->start_lock);
	145
	146	return NULL;
	147	}
	148
	149	/* calculate mean and standard deviation to complete an unit of work */
	150	static void calibration_stats(void)
	151	{
	152	int i;
680b5abc	153	double sum = 0.0, var = 0.0;
f2a2ce0e HL	154	struct idle_prof_thread *ipt;
	155
	156	for (i = 0; i < ipc.nr_cpus; i++) {
	157	ipt = &ipc.ipts[i];
	158	sum += ipt->cali_time;
	159	}
	160
	161	ipc.cali_mean = sum/ipc.nr_cpus;
	162
	163	for (i = 0; i < ipc.nr_cpus; i++) {
	164	ipt = &ipc.ipts[i];
	165	var += pow(ipt->cali_time-ipc.cali_mean, 2);
	166	}
	167
	168	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
	169	}
	170
	171	void fio_idle_prof_init(void)
	172	{
	173	int i, ret;
	174	struct timeval tp;
	175	struct timespec ts;
680b5abc	176	pthread_attr_t tattr;
f2a2ce0e HL	177	struct idle_prof_thread *ipt;
	178
	179	ipc.nr_cpus = cpus_online();
	180	ipc.status = IDLE_PROF_STATUS_OK;
	181
	182	if (ipc.opt == IDLE_PROF_OPT_NONE)
	183	return;
	184
	185	if ((ret = pthread_attr_init(&tattr))) {
	186	log_err("fio: pthread_attr_init %s\n", strerror(ret));
	187	return;
	188	}
	189	if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
	190	log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
	191	return;
	192	}
	193
	194	ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
	195	if (!ipc.ipts) {
	196	log_err("fio: malloc failed\n");
	197	return;
	198	}
	199
	200	ipc.buf = malloc(ipc.nr_cpus * page_size);
	201	if (!ipc.buf) {
	202	log_err("fio: malloc failed\n");
	203	free(ipc.ipts);
	204	return;
	205	}
	206
680b5abc JA	207	/*
680b5abc JA	208	* profiling aborts on any single thread failure since the
f2a2ce0e HL	209	* result won't be accurate if any cpu is not used.
	210	*/
	211	for (i = 0; i < ipc.nr_cpus; i++) {
	212	ipt = &ipc.ipts[i];
	213
	214	ipt->cpu = i;
	215	ipt->state = TD_NOT_CREATED;
	216	ipt->data = (unsigned char )(ipc.buf + page_size i);
	217
	218	if ((ret = pthread_mutex_init(&ipt->init_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_mutex_init(&ipt->start_lock, NULL))) {
	225	ipc.status = IDLE_PROF_STATUS_ABORT;
	226	log_err("fio: pthread_mutex_init %s\n", strerror(ret));
	227	break;
	228	}
	229
	230	if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
	231	ipc.status = IDLE_PROF_STATUS_ABORT;
	232	log_err("fio: pthread_cond_init %s\n", strerror(ret));
	233	break;
	234	}
	235
	236	/* make sure all threads are spawned before they start */
	237	pthread_mutex_lock(&ipt->init_lock);
	238
	239	/* make sure all threads finish init before profiling starts */
	240	pthread_mutex_lock(&ipt->start_lock);
	241
	242	if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
	243	ipc.status = IDLE_PROF_STATUS_ABORT;
	244	log_err("fio: pthread_create %s\n", strerror(ret));
	245	break;
680b5abc	246	} else
f2a2ce0e	247	ipt->state = TD_CREATED;
f2a2ce0e HL	248
	249	if ((ret = pthread_detach(ipt->thread))) {
	250	/* log error and let the thread spin */
	251	log_err("fio: pthread_detatch %s\n", strerror(ret));
	252	}
	253	}
	254
680b5abc JA	255	/*
	256	* let good threads continue so that they can exit
	257	* if errors on other threads occurred previously.
f2a2ce0e HL	258	*/
	259	for (i = 0; i < ipc.nr_cpus; i++) {
	260	ipt = &ipc.ipts[i];
	261	pthread_mutex_unlock(&ipt->init_lock);
	262	}
	263
	264	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	265	return;
	266
	267	/* wait for calibration to finish */
	268	for (i = 0; i < ipc.nr_cpus; i++) {
	269	ipt = &ipc.ipts[i];
	270	pthread_mutex_lock(&ipt->init_lock);
680b5abc JA	271	while ((ipt->state != TD_EXITED) &&
680b5abc JA	272	(ipt->state!=TD_INITIALIZED)) {
f2a2ce0e HL	273	fio_gettime(&tp, NULL);
	274	ts.tv_sec = tp.tv_sec + 1;
	275	ts.tv_nsec = tp.tv_usec * 1000;
	276	pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
	277	}
	278	pthread_mutex_unlock(&ipt->init_lock);
	279
680b5abc JA	280	/*
680b5abc JA	281	* any thread failed to initialize would abort other threads
f2a2ce0e HL	282	* later after fio_idle_prof_start.
	283	*/
	284	if (ipt->state == TD_EXITED)
	285	ipc.status = IDLE_PROF_STATUS_ABORT;
	286	}
	287
	288	if (ipc.status != IDLE_PROF_STATUS_ABORT)
	289	calibration_stats();
680b5abc	290	else
f2a2ce0e HL	291	ipc.cali_mean = ipc.cali_stddev = 0.0;
	292
	293	if (ipc.opt == IDLE_PROF_OPT_CALI)
	294	ipc.status = IDLE_PROF_STATUS_CALI_STOP;
	295	}
	296
	297	void fio_idle_prof_start(void)
	298	{
	299	int i;
	300	struct idle_prof_thread *ipt;
	301
	302	if (ipc.opt == IDLE_PROF_OPT_NONE)
	303	return;
	304
	305	/* unlock regardless abort is set or not */
	306	for (i = 0; i < ipc.nr_cpus; i++) {
	307	ipt = &ipc.ipts[i];
	308	pthread_mutex_unlock(&ipt->start_lock);
	309	}
	310	}
	311
	312	void fio_idle_prof_stop(void)
	313	{
	314	int i;
	315	uint64_t runt;
	316	struct timeval tp;
	317	struct timespec ts;
	318	struct idle_prof_thread *ipt;
	319
	320	if (ipc.opt == IDLE_PROF_OPT_NONE)
	321	return;
	322
	323	if (ipc.opt == IDLE_PROF_OPT_CALI)
	324	return;
	325
	326	ipc.status = IDLE_PROF_STATUS_PROF_STOP;
	327
	328	/* wait for all threads to exit from profiling */
	329	for (i = 0; i < ipc.nr_cpus; i++) {
	330	ipt = &ipc.ipts[i];
	331	pthread_mutex_lock(&ipt->start_lock);
680b5abc JA	332	while ((ipt->state != TD_EXITED) &&
680b5abc JA	333	(ipt->state!=TD_NOT_CREATED)) {
f2a2ce0e HL	334	fio_gettime(&tp, NULL);
	335	ts.tv_sec = tp.tv_sec + 1;
	336	ts.tv_nsec = tp.tv_usec * 1000;
	337	/* timed wait in case a signal is not received */
	338	pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
	339	}
	340	pthread_mutex_unlock(&ipt->start_lock);
	341
	342	/* calculate idleness */
	343	if (ipc.cali_mean != 0.0) {
	344	runt = utime_since(&ipt->tps, &ipt->tpe);
9da67e75 JA	345	if (runt)
	346	ipt->idleness = ipt->loops * ipc.cali_mean / runt;
	347	else
	348	ipt->idleness = 0.0;
680b5abc	349	} else
f2a2ce0e HL	350	ipt->idleness = 0.0;
	351	}
	352
680b5abc JA	353	/*
680b5abc JA	354	* memory allocations are freed via explicit fio_idle_prof_cleanup
f2a2ce0e HL	355	* after profiling stats are collected by apps.
f2a2ce0e HL	356	*/
f2a2ce0e HL	357	}
f2a2ce0e HL	358
680b5abc JA	359	/*
680b5abc JA	360	* return system idle percentage when cpu is -1;
f2a2ce0e HL	361	* return one cpu idle percentage otherwise.
	362	*/
	363	static double fio_idle_prof_cpu_stat(int cpu)
	364	{
	365	int i, nr_cpus = ipc.nr_cpus;
	366	struct idle_prof_thread *ipt;
	367	double p = 0.0;
	368
	369	if (ipc.opt == IDLE_PROF_OPT_NONE)
	370	return 0.0;
	371
	372	if ((cpu >= nr_cpus) \|\| (cpu < -1)) {
	373	log_err("fio: idle profiling invalid cpu index\n");
	374	return 0.0;
	375	}
	376
	377	if (cpu == -1) {
	378	for (i = 0; i < nr_cpus; i++) {
	379	ipt = &ipc.ipts[i];
	380	p += ipt->idleness;
	381	}
	382	p /= nr_cpus;
	383	} else {
	384	ipt = &ipc.ipts[cpu];
	385	p = ipt->idleness;
	386	}
	387
680b5abc	388	return p * 100.0;
f2a2ce0e HL	389	}
f2a2ce0e HL	390
10aa136b	391	static void fio_idle_prof_cleanup(void)
f2a2ce0e HL	392	{
	393	if (ipc.ipts) {
	394	free(ipc.ipts);
	395	ipc.ipts = NULL;
	396	}
	397
	398	if (ipc.buf) {
	399	free(ipc.buf);
	400	ipc.buf = NULL;
	401	}
	402	}
	403
	404	int fio_idle_prof_parse_opt(const char *args)
	405	{
	406	ipc.opt = IDLE_PROF_OPT_NONE; /* default */
	407
	408	if (!args) {
	409	log_err("fio: empty idle-prof option string\n");
	410	return -1;
	411	}
	412
7e09a9f1	413	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	414	if (strcmp("calibrate", args) == 0) {
	415	ipc.opt = IDLE_PROF_OPT_CALI;
	416	fio_idle_prof_init();
	417	fio_idle_prof_start();
	418	fio_idle_prof_stop();
	419	show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL);
	420	return 1;
	421	} else if (strcmp("system", args) == 0) {
	422	ipc.opt = IDLE_PROF_OPT_SYSTEM;
	423	return 0;
	424	} else if (strcmp("percpu", args) == 0) {
	425	ipc.opt = IDLE_PROF_OPT_PERCPU;
	426	return 0;
	427	} else {
4e0a8fa2	428	log_err("fio: incorrect idle-prof option: %s\n", args);
f2a2ce0e HL	429	return -1;
	430	}
	431	#else
	432	log_err("fio: idle-prof not supported on this platform\n");
	433	return -1;
	434	#endif
	435	}
	436
	437	void show_idle_prof_stats(int output, struct json_object *parent)
	438	{
	439	int i, nr_cpus = ipc.nr_cpus;
	440	struct json_object *tmp;
	441	char s[MAX_CPU_STR_LEN];
680b5abc	442
f2a2ce0e HL	443	if (output == FIO_OUTPUT_NORMAL) {
	444	if (ipc.opt > IDLE_PROF_OPT_CALI)
	445	log_info("\nCPU idleness:\n");
	446	else if (ipc.opt == IDLE_PROF_OPT_CALI)
	447	log_info("CPU idleness:\n");
	448
	449	if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
	450	log_info(" system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
	451
	452	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
	453	log_info(" percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
680b5abc	454	for (i = 1; i < nr_cpus; i++)
f2a2ce0e	455	log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i));
f2a2ce0e HL	456	log_info("\n");
	457	}
	458
	459	if (ipc.opt >= IDLE_PROF_OPT_CALI) {
	460	log_info(" unit work: mean=%3.2fus,", ipc.cali_mean);
	461	log_info(" stddev=%3.2f\n", ipc.cali_stddev);
	462	}
	463
	464	/* dynamic mem allocations can now be freed */
	465	if (ipc.opt != IDLE_PROF_OPT_NONE)
	466	fio_idle_prof_cleanup();
	467
	468	return;
	469	}
680b5abc	470
f2a2ce0e HL	471	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) {
	472	if (!parent)
	473	return;
	474
	475	tmp = json_create_object();
	476	if (!tmp)
	477	return;
	478
	479	json_object_add_value_object(parent, "cpu_idleness", tmp);
	480	json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
	481
	482	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
680b5abc	483	for (i = 0; i < nr_cpus; i++) {
f2a2ce0e HL	484	snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
	485	json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
	486	}
	487	}
	488
	489	json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
	490	json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
	491
	492	fio_idle_prof_cleanup();
f2a2ce0e HL	493	}
f2a2ce0e HL	494	}