[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) {
		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 */
	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 */
735ed278 JA	112	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
735ed278 JA	113	pthread_mutex_unlock(&ipt->start_lock);
f2a2ce0e	114	return NULL;
735ed278	115	}
f2a2ce0e HL	116
f2a2ce0e HL	117	/* exit if we are doing calibration only */
735ed278 JA	118	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
735ed278 JA	119	pthread_mutex_unlock(&ipt->start_lock);
f2a2ce0e	120	return NULL;
735ed278	121	}
f2a2ce0e HL	122
	123	fio_gettime(&ipt->tps, NULL);
	124	ipt->state = TD_RUNNING;
	125
	126	j = 0;
	127	while (1) {
680b5abc JA	128	for (k = 0; k < page_size; k++) {
680b5abc JA	129	ipt->data[(k + j) % page_size] = k % 256;
f2a2ce0e HL	130	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
	131	fio_gettime(&ipt->tpe, NULL);
	132	goto idle_prof_done;
	133	}
	134	}
	135	j++;
	136	}
	137
	138	idle_prof_done:
	139
680b5abc	140	ipt->loops = j + (double) k / page_size;
f2a2ce0e HL	141	ipt->state = TD_EXITED;
	142	pthread_mutex_unlock(&ipt->start_lock);
	143
	144	return NULL;
	145	}
	146
	147	/* calculate mean and standard deviation to complete an unit of work */
	148	static void calibration_stats(void)
	149	{
	150	int i;
680b5abc	151	double sum = 0.0, var = 0.0;
f2a2ce0e HL	152	struct idle_prof_thread *ipt;
	153
	154	for (i = 0; i < ipc.nr_cpus; i++) {
	155	ipt = &ipc.ipts[i];
	156	sum += ipt->cali_time;
	157	}
	158
	159	ipc.cali_mean = sum/ipc.nr_cpus;
	160
	161	for (i = 0; i < ipc.nr_cpus; i++) {
	162	ipt = &ipc.ipts[i];
	163	var += pow(ipt->cali_time-ipc.cali_mean, 2);
	164	}
	165
	166	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
	167	}
	168
	169	void fio_idle_prof_init(void)
	170	{
	171	int i, ret;
	172	struct timeval tp;
	173	struct timespec ts;
680b5abc	174	pthread_attr_t tattr;
f2a2ce0e HL	175	struct idle_prof_thread *ipt;
	176
	177	ipc.nr_cpus = cpus_online();
	178	ipc.status = IDLE_PROF_STATUS_OK;
	179
	180	if (ipc.opt == IDLE_PROF_OPT_NONE)
	181	return;
	182
	183	if ((ret = pthread_attr_init(&tattr))) {
	184	log_err("fio: pthread_attr_init %s\n", strerror(ret));
	185	return;
	186	}
	187	if ((ret = pthread_attr_setscope(&tattr, PTHREAD_SCOPE_SYSTEM))) {
	188	log_err("fio: pthread_attr_setscope %s\n", strerror(ret));
	189	return;
	190	}
	191
	192	ipc.ipts = malloc(ipc.nr_cpus * sizeof(struct idle_prof_thread));
	193	if (!ipc.ipts) {
	194	log_err("fio: malloc failed\n");
	195	return;
	196	}
	197
	198	ipc.buf = malloc(ipc.nr_cpus * page_size);
	199	if (!ipc.buf) {
	200	log_err("fio: malloc failed\n");
	201	free(ipc.ipts);
	202	return;
	203	}
	204
680b5abc JA	205	/*
680b5abc JA	206	* profiling aborts on any single thread failure since the
f2a2ce0e HL	207	* result won't be accurate if any cpu is not used.
	208	*/
	209	for (i = 0; i < ipc.nr_cpus; i++) {
	210	ipt = &ipc.ipts[i];
	211
	212	ipt->cpu = i;
	213	ipt->state = TD_NOT_CREATED;
	214	ipt->data = (unsigned char )(ipc.buf + page_size i);
	215
	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));
	219	break;
	220	}
	221
	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));
	225	break;
	226	}
	227
	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));
	231	break;
	232	}
	233
	234	/* make sure all threads are spawned before they start */
	235	pthread_mutex_lock(&ipt->init_lock);
	236
	237	/* make sure all threads finish init before profiling starts */
	238	pthread_mutex_lock(&ipt->start_lock);
	239
	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));
	243	break;
680b5abc	244	} else
f2a2ce0e	245	ipt->state = TD_CREATED;
f2a2ce0e HL	246
	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));
	250	}
	251	}
	252
680b5abc JA	253	/*
	254	* let good threads continue so that they can exit
	255	* if errors on other threads occurred previously.
f2a2ce0e HL	256	*/
	257	for (i = 0; i < ipc.nr_cpus; i++) {
	258	ipt = &ipc.ipts[i];
	259	pthread_mutex_unlock(&ipt->init_lock);
	260	}
	261
	262	if (ipc.status == IDLE_PROF_STATUS_ABORT)
	263	return;
	264
	265	/* wait for calibration to finish */
	266	for (i = 0; i < ipc.nr_cpus; i++) {
	267	ipt = &ipc.ipts[i];
	268	pthread_mutex_lock(&ipt->init_lock);
680b5abc JA	269	while ((ipt->state != TD_EXITED) &&
680b5abc JA	270	(ipt->state!=TD_INITIALIZED)) {
f2a2ce0e HL	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);
	275	}
	276	pthread_mutex_unlock(&ipt->init_lock);
	277
680b5abc JA	278	/*
680b5abc JA	279	* any thread failed to initialize would abort other threads
f2a2ce0e HL	280	* later after fio_idle_prof_start.
	281	*/
	282	if (ipt->state == TD_EXITED)
	283	ipc.status = IDLE_PROF_STATUS_ABORT;
	284	}
	285
	286	if (ipc.status != IDLE_PROF_STATUS_ABORT)
	287	calibration_stats();
680b5abc	288	else
f2a2ce0e HL	289	ipc.cali_mean = ipc.cali_stddev = 0.0;
	290
	291	if (ipc.opt == IDLE_PROF_OPT_CALI)
	292	ipc.status = IDLE_PROF_STATUS_CALI_STOP;
	293	}
	294
	295	void fio_idle_prof_start(void)
	296	{
	297	int i;
	298	struct idle_prof_thread *ipt;
	299
	300	if (ipc.opt == IDLE_PROF_OPT_NONE)
	301	return;
	302
	303	/* unlock regardless abort is set or not */
	304	for (i = 0; i < ipc.nr_cpus; i++) {
	305	ipt = &ipc.ipts[i];
	306	pthread_mutex_unlock(&ipt->start_lock);
	307	}
	308	}
	309
	310	void fio_idle_prof_stop(void)
	311	{
	312	int i;
	313	uint64_t runt;
	314	struct timeval tp;
	315	struct timespec ts;
	316	struct idle_prof_thread *ipt;
	317
	318	if (ipc.opt == IDLE_PROF_OPT_NONE)
	319	return;
	320
	321	if (ipc.opt == IDLE_PROF_OPT_CALI)
	322	return;
	323
	324	ipc.status = IDLE_PROF_STATUS_PROF_STOP;
	325
	326	/* wait for all threads to exit from profiling */
	327	for (i = 0; i < ipc.nr_cpus; i++) {
	328	ipt = &ipc.ipts[i];
	329	pthread_mutex_lock(&ipt->start_lock);
680b5abc JA	330	while ((ipt->state != TD_EXITED) &&
680b5abc JA	331	(ipt->state!=TD_NOT_CREATED)) {
f2a2ce0e HL	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);
	337	}
	338	pthread_mutex_unlock(&ipt->start_lock);
	339
	340	/* calculate idleness */
	341	if (ipc.cali_mean != 0.0) {
	342	runt = utime_since(&ipt->tps, &ipt->tpe);
9da67e75 JA	343	if (runt)
	344	ipt->idleness = ipt->loops * ipc.cali_mean / runt;
	345	else
	346	ipt->idleness = 0.0;
680b5abc	347	} else
f2a2ce0e HL	348	ipt->idleness = 0.0;
	349	}
	350
680b5abc JA	351	/*
680b5abc JA	352	* memory allocations are freed via explicit fio_idle_prof_cleanup
f2a2ce0e HL	353	* after profiling stats are collected by apps.
f2a2ce0e HL	354	*/
f2a2ce0e HL	355	}
f2a2ce0e HL	356
680b5abc JA	357	/*
680b5abc JA	358	* return system idle percentage when cpu is -1;
f2a2ce0e HL	359	* return one cpu idle percentage otherwise.
	360	*/
	361	static double fio_idle_prof_cpu_stat(int cpu)
	362	{
	363	int i, nr_cpus = ipc.nr_cpus;
	364	struct idle_prof_thread *ipt;
	365	double p = 0.0;
	366
	367	if (ipc.opt == IDLE_PROF_OPT_NONE)
	368	return 0.0;
	369
	370	if ((cpu >= nr_cpus) \|\| (cpu < -1)) {
	371	log_err("fio: idle profiling invalid cpu index\n");
	372	return 0.0;
	373	}
	374
	375	if (cpu == -1) {
	376	for (i = 0; i < nr_cpus; i++) {
	377	ipt = &ipc.ipts[i];
	378	p += ipt->idleness;
	379	}
	380	p /= nr_cpus;
	381	} else {
	382	ipt = &ipc.ipts[cpu];
	383	p = ipt->idleness;
	384	}
	385
680b5abc	386	return p * 100.0;
f2a2ce0e HL	387	}
f2a2ce0e HL	388
10aa136b	389	static void fio_idle_prof_cleanup(void)
f2a2ce0e HL	390	{
	391	if (ipc.ipts) {
	392	free(ipc.ipts);
	393	ipc.ipts = NULL;
	394	}
	395
	396	if (ipc.buf) {
	397	free(ipc.buf);
	398	ipc.buf = NULL;
	399	}
	400	}
	401
	402	int fio_idle_prof_parse_opt(const char *args)
	403	{
	404	ipc.opt = IDLE_PROF_OPT_NONE; /* default */
	405
	406	if (!args) {
	407	log_err("fio: empty idle-prof option string\n");
	408	return -1;
	409	}
	410
7e09a9f1	411	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	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);
	418	return 1;
	419	} else if (strcmp("system", args) == 0) {
	420	ipc.opt = IDLE_PROF_OPT_SYSTEM;
	421	return 0;
	422	} else if (strcmp("percpu", args) == 0) {
	423	ipc.opt = IDLE_PROF_OPT_PERCPU;
	424	return 0;
	425	} else {
4e0a8fa2	426	log_err("fio: incorrect idle-prof option: %s\n", args);
f2a2ce0e HL	427	return -1;
	428	}
	429	#else
	430	log_err("fio: idle-prof not supported on this platform\n");
	431	return -1;
	432	#endif
	433	}
	434
	435	void show_idle_prof_stats(int output, struct json_object *parent)
	436	{
	437	int i, nr_cpus = ipc.nr_cpus;
	438	struct json_object *tmp;
	439	char s[MAX_CPU_STR_LEN];
680b5abc	440
f2a2ce0e HL	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");
	446
	447	if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
	448	log_info(" system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
	449
	450	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
	451	log_info(" percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
680b5abc	452	for (i = 1; i < nr_cpus; i++)
f2a2ce0e	453	log_info(", %3.2f%%", fio_idle_prof_cpu_stat(i));
f2a2ce0e HL	454	log_info("\n");
	455	}
	456
	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);
	460	}
	461
	462	/* dynamic mem allocations can now be freed */
	463	if (ipc.opt != IDLE_PROF_OPT_NONE)
	464	fio_idle_prof_cleanup();
	465
	466	return;
	467	}
680b5abc	468
f2a2ce0e HL	469	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output == FIO_OUTPUT_JSON)) {
	470	if (!parent)
	471	return;
	472
	473	tmp = json_create_object();
	474	if (!tmp)
	475	return;
	476
	477	json_object_add_value_object(parent, "cpu_idleness", tmp);
	478	json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
	479
	480	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
680b5abc	481	for (i = 0; i < nr_cpus; i++) {
f2a2ce0e HL	482	snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
	483	json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
	484	}
	485	}
	486
	487	json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
	488	json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
	489
	490	fio_idle_prof_cleanup();
f2a2ce0e HL	491	}
f2a2ce0e HL	492	}