[fio.git] / idletime.c

#include <math.h>
#include "fio.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 timespec 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 void free_cpu_affinity(struct idle_prof_thread *ipt)
{
#if defined(FIO_HAVE_CPU_AFFINITY)
	fio_cpuset_exit(&ipt->cpu_mask);
#endif
}

static int set_cpu_affinity(struct idle_prof_thread *ipt)
{
#if defined(FIO_HAVE_CPU_AFFINITY)
	if (fio_cpuset_init(&ipt->cpu_mask)) {
		log_err("fio: cpuset init failed\n");
		return -1;
	}

	fio_cpu_set(&ipt->cpu_mask, ipt->cpu);

	if (fio_setaffinity(gettid(), ipt->cpu_mask)) {
		log_err("fio: fio_setaffinity failed\n");
		fio_cpuset_exit(&ipt->cpu_mask);
		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);
		goto do_exit;
	}

	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);
		goto do_exit;
	}

	/* exit if we are doing calibration only */
	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
		pthread_mutex_unlock(&ipt->start_lock);
		goto do_exit;
	}

	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);

do_exit:
	free_cpu_affinity(ipt);
	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 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_detach %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(&ts, NULL);
			ts.tv_sec += 1;
			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 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(&ts, NULL);
			ts.tv_sec += 1;
			/* 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;
}

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, 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,
			  struct buf_output *out)
{
	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_buf(out, "\nCPU idleness:\n");
		else if (ipc.opt == IDLE_PROF_OPT_CALI)
			log_buf(out, "CPU idleness:\n");

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

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

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

		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);
	}
}
Commit	Line	Data
f2a2ce0e	1	#include <math.h>
3d2d14bc	2	#include "fio.h"
f2a2ce0e HL	3	#include "json.h"
	4	#include "idletime.h"
	5
	6	static volatile struct idle_prof_common ipc;
	7
680b5abc JA	8	/*
680b5abc JA	9	* Get time to complete an unit work on a particular cpu.
f2a2ce0e HL	10	* The minimum number in CALIBRATE_RUNS runs is returned.
	11	*/
	12	static double calibrate_unit(unsigned char *data)
	13	{
	14	unsigned long t, i, j, k;
8b6a404c	15	struct timespec tps;
f2a2ce0e HL	16	double tunit = 0.0;
f2a2ce0e HL	17
680b5abc	18	for (i = 0; i < CALIBRATE_RUNS; i++) {
f2a2ce0e HL	19
	20	fio_gettime(&tps, NULL);
	21	/* scale for less variance */
680b5abc	22	for (j = 0; j < CALIBRATE_SCALE; j++) {
f2a2ce0e HL	23	/* unit of work */
f2a2ce0e HL	24	for (k=0; k < page_size; k++) {
680b5abc JA	25	data[(k + j) % page_size] = k % 256;
	26	/*
	27	* we won't see STOP here. this is to match
f2a2ce0e HL	28	* the same statement in the profiling loop.
	29	*/
	30	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP)
	31	return 0.0;
	32	}
	33	}
	34
	35	t = utime_since_now(&tps);
	36	if (!t)
	37	continue;
	38
	39	/* get the minimum time to complete CALIBRATE_SCALE units */
680b5abc	40	if ((i == 0) \|\| ((double)t < tunit))
f2a2ce0e HL	41	tunit = (double)t;
	42	}
	43
680b5abc	44	return tunit / CALIBRATE_SCALE;
f2a2ce0e HL	45	}
f2a2ce0e HL	46
54ed125b JA	47	static void free_cpu_affinity(struct idle_prof_thread *ipt)
	48	{
	49	#if defined(FIO_HAVE_CPU_AFFINITY)
	50	fio_cpuset_exit(&ipt->cpu_mask);
	51	#endif
	52	}
	53
59358c8e JA	54	static int set_cpu_affinity(struct idle_prof_thread *ipt)
	55	{
	56	#if defined(FIO_HAVE_CPU_AFFINITY)
54ed125b JA	57	if (fio_cpuset_init(&ipt->cpu_mask)) {
	58	log_err("fio: cpuset init failed\n");
	59	return -1;
	60	}
59358c8e	61
54ed125b	62	fio_cpu_set(&ipt->cpu_mask, ipt->cpu);
59358c8e	63
54ed125b	64	if (fio_setaffinity(gettid(), ipt->cpu_mask)) {
59358c8e	65	log_err("fio: fio_setaffinity failed\n");
54ed125b	66	fio_cpuset_exit(&ipt->cpu_mask);
59358c8e JA	67	return -1;
	68	}
	69
	70	return 0;
	71	#else
	72	log_err("fio: fio_setaffinity not supported\n");
	73	return -1;
	74	#endif
	75	}
	76
f2a2ce0e HL	77	static void idle_prof_thread_fn(void data)
	78	{
	79	int retval;
	80	unsigned long j, k;
	81	struct idle_prof_thread *ipt = data;
	82
	83	/* wait for all threads are spawned */
	84	pthread_mutex_lock(&ipt->init_lock);
	85
	86	/* exit if any other thread failed to start */
d5b351d4 JA	87	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
d5b351d4 JA	88	pthread_mutex_unlock(&ipt->init_lock);
f2a2ce0e	89	return NULL;
d5b351d4	90	}
f2a2ce0e	91
59358c8e	92	retval = set_cpu_affinity(ipt);
f2a2ce0e HL	93	if (retval == -1) {
	94	ipt->state = TD_EXITED;
	95	pthread_mutex_unlock(&ipt->init_lock);
	96	return NULL;
	97	}
	98
	99	ipt->cali_time = calibrate_unit(ipt->data);
	100
	101	/* delay to set IDLE class till now for better calibration accuracy */
7e09a9f1	102	#if defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	103	if ((retval = fio_set_sched_idle()))
	104	log_err("fio: fio_set_sched_idle failed\n");
	105	#else
	106	retval = -1;
	107	log_err("fio: fio_set_sched_idle not supported\n");
	108	#endif
	109	if (retval == -1) {
	110	ipt->state = TD_EXITED;
	111	pthread_mutex_unlock(&ipt->init_lock);
54ed125b	112	goto do_exit;
f2a2ce0e HL	113	}
	114
	115	ipt->state = TD_INITIALIZED;
	116
	117	/* signal the main thread that calibration is done */
	118	pthread_cond_signal(&ipt->cond);
	119	pthread_mutex_unlock(&ipt->init_lock);
	120
	121	/* wait for other calibration to finish */
	122	pthread_mutex_lock(&ipt->start_lock);
	123
	124	/* exit if other threads failed to initialize */
735ed278 JA	125	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
735ed278 JA	126	pthread_mutex_unlock(&ipt->start_lock);
54ed125b	127	goto do_exit;
735ed278	128	}
f2a2ce0e HL	129
f2a2ce0e HL	130	/* exit if we are doing calibration only */
735ed278 JA	131	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
735ed278 JA	132	pthread_mutex_unlock(&ipt->start_lock);
54ed125b	133	goto do_exit;
735ed278	134	}
f2a2ce0e HL	135
	136	fio_gettime(&ipt->tps, NULL);
	137	ipt->state = TD_RUNNING;
	138
	139	j = 0;
	140	while (1) {
680b5abc JA	141	for (k = 0; k < page_size; k++) {
680b5abc JA	142	ipt->data[(k + j) % page_size] = k % 256;
f2a2ce0e HL	143	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
	144	fio_gettime(&ipt->tpe, NULL);
	145	goto idle_prof_done;
	146	}
	147	}
	148	j++;
	149	}
	150
	151	idle_prof_done:
	152
680b5abc	153	ipt->loops = j + (double) k / page_size;
f2a2ce0e HL	154	ipt->state = TD_EXITED;
	155	pthread_mutex_unlock(&ipt->start_lock);
	156
54ed125b JA	157	do_exit:
54ed125b JA	158	free_cpu_affinity(ipt);
f2a2ce0e HL	159	return NULL;
	160	}
	161
	162	/* calculate mean and standard deviation to complete an unit of work */
	163	static void calibration_stats(void)
	164	{
	165	int i;
680b5abc	166	double sum = 0.0, var = 0.0;
f2a2ce0e HL	167	struct idle_prof_thread *ipt;
	168
	169	for (i = 0; i < ipc.nr_cpus; i++) {
	170	ipt = &ipc.ipts[i];
	171	sum += ipt->cali_time;
	172	}
	173
	174	ipc.cali_mean = sum/ipc.nr_cpus;
	175
	176	for (i = 0; i < ipc.nr_cpus; i++) {
	177	ipt = &ipc.ipts[i];
	178	var += pow(ipt->cali_time-ipc.cali_mean, 2);
	179	}
	180
	181	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
	182	}
	183
	184	void fio_idle_prof_init(void)
	185	{
	186	int i, ret;
f2a2ce0e	187	struct timespec ts;
680b5abc	188	pthread_attr_t tattr;
f2a2ce0e HL	189	struct idle_prof_thread *ipt;
	190
	191	ipc.nr_cpus = cpus_online();
	192	ipc.status = IDLE_PROF_STATUS_OK;
	193
	194	if (ipc.opt == IDLE_PROF_OPT_NONE)
	195	return;
	196
	197	if ((ret = pthread_attr_init(&tattr))) {
	198	log_err("fio: pthread_attr_init %s\n", strerror(ret));
	199	return;
	200	}
	201	if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
	202	log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
	203	return;
	204	}
	205
	206	ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
	207	if (!ipc.ipts) {
	208	log_err("fio: malloc failed\n");
	209	return;
	210	}
	211
	212	ipc.buf = malloc(ipc.nr_cpus * page_size);
	213	if (!ipc.buf) {
	214	log_err("fio: malloc failed\n");
	215	free(ipc.ipts);
	216	return;
	217	}
	218
680b5abc JA	219	/*
680b5abc JA	220	* profiling aborts on any single thread failure since the
f2a2ce0e HL	221	* result won't be accurate if any cpu is not used.
	222	*/
	223	for (i = 0; i < ipc.nr_cpus; i++) {
	224	ipt = &ipc.ipts[i];
	225
	226	ipt->cpu = i;
	227	ipt->state = TD_NOT_CREATED;
	228	ipt->data = (unsigned char )(ipc.buf + page_size i);
	229
	230	if ((ret = pthread_mutex_init(&ipt->init_lock, NULL))) {
	231	ipc.status = IDLE_PROF_STATUS_ABORT;
	232	log_err("fio: pthread_mutex_init %s\n", strerror(ret));
	233	break;
	234	}
	235
	236	if ((ret = pthread_mutex_init(&ipt->start_lock, NULL))) {
	237	ipc.status = IDLE_PROF_STATUS_ABORT;
	238	log_err("fio: pthread_mutex_init %s\n", strerror(ret));
	239	break;
	240	}
	241
	242	if ((ret = pthread_cond_init(&ipt->cond, NULL))) {
	243	ipc.status = IDLE_PROF_STATUS_ABORT;
	244	log_err("fio: pthread_cond_init %s\n", strerror(ret));
	245	break;
	246	}
	247
	248	/* make sure all threads are spawned before they start */
	249	pthread_mutex_lock(&ipt->init_lock);
	250
	251	/* make sure all threads finish init before profiling starts */
	252	pthread_mutex_lock(&ipt->start_lock);
	253
	254	if ((ret = pthread_create(&ipt->thread, &tattr, idle_prof_thread_fn, ipt))) {
	255	ipc.status = IDLE_PROF_STATUS_ABORT;
	256	log_err("fio: pthread_create %s\n", strerror(ret));
	257	break;
680b5abc	258	} else
f2a2ce0e	259	ipt->state = TD_CREATED;
f2a2ce0e HL	260
	261	if ((ret = pthread_detach(ipt->thread))) {
	262	/* log error and let the thread spin */
5bb39b7c	263	log_err("fio: pthread_detach %s\n", strerror(ret));
f2a2ce0e HL	264	}
	265	}
	266
680b5abc JA	267	/*
	268	* let good threads continue so that they can exit
	269	* if errors on other threads occurred previously.
f2a2ce0e HL	270	*/
	271	for (i = 0; i < ipc.nr_cpus; i++) {
	272	ipt = &ipc.ipts[i];
	273	pthread_mutex_unlock(&ipt->init_lock);
	274	}
	275
	276	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	277	return;
	278
	279	/* wait for calibration to finish */
	280	for (i = 0; i < ipc.nr_cpus; i++) {
	281	ipt = &ipc.ipts[i];
	282	pthread_mutex_lock(&ipt->init_lock);
680b5abc JA	283	while ((ipt->state != TD_EXITED) &&
680b5abc JA	284	(ipt->state!=TD_INITIALIZED)) {
8b6a404c VF	285	fio_gettime(&ts, NULL);
8b6a404c VF	286	ts.tv_sec += 1;
f2a2ce0e HL	287	pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
	288	}
	289	pthread_mutex_unlock(&ipt->init_lock);
	290
680b5abc JA	291	/*
680b5abc JA	292	* any thread failed to initialize would abort other threads
f2a2ce0e HL	293	* later after fio_idle_prof_start.
	294	*/
	295	if (ipt->state == TD_EXITED)
	296	ipc.status = IDLE_PROF_STATUS_ABORT;
	297	}
	298
	299	if (ipc.status != IDLE_PROF_STATUS_ABORT)
	300	calibration_stats();
680b5abc	301	else
f2a2ce0e HL	302	ipc.cali_mean = ipc.cali_stddev = 0.0;
	303
	304	if (ipc.opt == IDLE_PROF_OPT_CALI)
	305	ipc.status = IDLE_PROF_STATUS_CALI_STOP;
	306	}
	307
	308	void fio_idle_prof_start(void)
	309	{
	310	int i;
	311	struct idle_prof_thread *ipt;
	312
	313	if (ipc.opt == IDLE_PROF_OPT_NONE)
	314	return;
	315
	316	/* unlock regardless abort is set or not */
	317	for (i = 0; i < ipc.nr_cpus; i++) {
	318	ipt = &ipc.ipts[i];
	319	pthread_mutex_unlock(&ipt->start_lock);
	320	}
	321	}
	322
	323	void fio_idle_prof_stop(void)
	324	{
	325	int i;
	326	uint64_t runt;
f2a2ce0e HL	327	struct timespec ts;
	328	struct idle_prof_thread *ipt;
	329
	330	if (ipc.opt == IDLE_PROF_OPT_NONE)
	331	return;
	332
	333	if (ipc.opt == IDLE_PROF_OPT_CALI)
	334	return;
	335
	336	ipc.status = IDLE_PROF_STATUS_PROF_STOP;
	337
	338	/* wait for all threads to exit from profiling */
	339	for (i = 0; i < ipc.nr_cpus; i++) {
	340	ipt = &ipc.ipts[i];
	341	pthread_mutex_lock(&ipt->start_lock);
680b5abc JA	342	while ((ipt->state != TD_EXITED) &&
680b5abc JA	343	(ipt->state!=TD_NOT_CREATED)) {
8b6a404c VF	344	fio_gettime(&ts, NULL);
8b6a404c VF	345	ts.tv_sec += 1;
f2a2ce0e HL	346	/* timed wait in case a signal is not received */
	347	pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
	348	}
	349	pthread_mutex_unlock(&ipt->start_lock);
	350
	351	/* calculate idleness */
	352	if (ipc.cali_mean != 0.0) {
	353	runt = utime_since(&ipt->tps, &ipt->tpe);
9da67e75 JA	354	if (runt)
	355	ipt->idleness = ipt->loops * ipc.cali_mean / runt;
	356	else
	357	ipt->idleness = 0.0;
680b5abc	358	} else
f2a2ce0e HL	359	ipt->idleness = 0.0;
	360	}
	361
680b5abc JA	362	/*
680b5abc JA	363	* memory allocations are freed via explicit fio_idle_prof_cleanup
f2a2ce0e HL	364	* after profiling stats are collected by apps.
f2a2ce0e HL	365	*/
f2a2ce0e HL	366	}
f2a2ce0e HL	367
680b5abc JA	368	/*
680b5abc JA	369	* return system idle percentage when cpu is -1;
f2a2ce0e HL	370	* return one cpu idle percentage otherwise.
	371	*/
	372	static double fio_idle_prof_cpu_stat(int cpu)
	373	{
	374	int i, nr_cpus = ipc.nr_cpus;
	375	struct idle_prof_thread *ipt;
	376	double p = 0.0;
	377
	378	if (ipc.opt == IDLE_PROF_OPT_NONE)
	379	return 0.0;
	380
	381	if ((cpu >= nr_cpus) \|\| (cpu < -1)) {
	382	log_err("fio: idle profiling invalid cpu index\n");
	383	return 0.0;
	384	}
	385
	386	if (cpu == -1) {
	387	for (i = 0; i < nr_cpus; i++) {
	388	ipt = &ipc.ipts[i];
	389	p += ipt->idleness;
	390	}
	391	p /= nr_cpus;
	392	} else {
	393	ipt = &ipc.ipts[cpu];
	394	p = ipt->idleness;
	395	}
	396
680b5abc	397	return p * 100.0;
f2a2ce0e HL	398	}
f2a2ce0e HL	399
7ceefc09	400	void fio_idle_prof_cleanup(void)
f2a2ce0e HL	401	{
	402	if (ipc.ipts) {
	403	free(ipc.ipts);
	404	ipc.ipts = NULL;
	405	}
	406
	407	if (ipc.buf) {
	408	free(ipc.buf);
	409	ipc.buf = NULL;
	410	}
	411	}
	412
	413	int fio_idle_prof_parse_opt(const char *args)
	414	{
	415	ipc.opt = IDLE_PROF_OPT_NONE; /* default */
	416
	417	if (!args) {
	418	log_err("fio: empty idle-prof option string\n");
	419	return -1;
	420	}
	421
7e09a9f1	422	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	423	if (strcmp("calibrate", args) == 0) {
	424	ipc.opt = IDLE_PROF_OPT_CALI;
	425	fio_idle_prof_init();
	426	fio_idle_prof_start();
	427	fio_idle_prof_stop();
a666cab8	428	show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, NULL);
f2a2ce0e HL	429	return 1;
	430	} else if (strcmp("system", args) == 0) {
	431	ipc.opt = IDLE_PROF_OPT_SYSTEM;
	432	return 0;
	433	} else if (strcmp("percpu", args) == 0) {
	434	ipc.opt = IDLE_PROF_OPT_PERCPU;
	435	return 0;
	436	} else {
4e0a8fa2	437	log_err("fio: incorrect idle-prof option: %s\n", args);
f2a2ce0e HL	438	return -1;
	439	}
	440	#else
	441	log_err("fio: idle-prof not supported on this platform\n");
	442	return -1;
	443	#endif
	444	}
	445
a666cab8 JA	446	void show_idle_prof_stats(int output, struct json_object *parent,
a666cab8 JA	447	struct buf_output *out)
f2a2ce0e HL	448	{
	449	int i, nr_cpus = ipc.nr_cpus;
	450	struct json_object *tmp;
	451	char s[MAX_CPU_STR_LEN];
680b5abc	452
f2a2ce0e HL	453	if (output == FIO_OUTPUT_NORMAL) {
f2a2ce0e HL	454	if (ipc.opt > IDLE_PROF_OPT_CALI)
a666cab8	455	log_buf(out, "\nCPU idleness:\n");
f2a2ce0e	456	else if (ipc.opt == IDLE_PROF_OPT_CALI)
a666cab8	457	log_buf(out, "CPU idleness:\n");
f2a2ce0e HL	458
f2a2ce0e HL	459	if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
a666cab8	460	log_buf(out, " system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
f2a2ce0e HL	461
f2a2ce0e HL	462	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
a666cab8	463	log_buf(out, " percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
680b5abc	464	for (i = 1; i < nr_cpus; i++)
a666cab8 JA	465	log_buf(out, ", %3.2f%%", fio_idle_prof_cpu_stat(i));
a666cab8 JA	466	log_buf(out, "\n");
f2a2ce0e HL	467	}
	468
	469	if (ipc.opt >= IDLE_PROF_OPT_CALI) {
a666cab8 JA	470	log_buf(out, " unit work: mean=%3.2fus,", ipc.cali_mean);
a666cab8 JA	471	log_buf(out, " stddev=%3.2f\n", ipc.cali_stddev);
f2a2ce0e HL	472	}
f2a2ce0e HL	473
f2a2ce0e HL	474	return;
f2a2ce0e HL	475	}
680b5abc	476
236d24df	477	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output & FIO_OUTPUT_JSON)) {
f2a2ce0e HL	478	if (!parent)
	479	return;
	480
	481	tmp = json_create_object();
	482	if (!tmp)
	483	return;
	484
	485	json_object_add_value_object(parent, "cpu_idleness", tmp);
	486	json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
	487
	488	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
680b5abc	489	for (i = 0; i < nr_cpus; i++) {
f2a2ce0e HL	490	snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
	491	json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
	492	}
	493	}
	494
	495	json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
	496	json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
f2a2ce0e HL	497	}
f2a2ce0e HL	498	}