[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 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 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, 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);
		}

		/* 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
54ed125b JA	46	static void free_cpu_affinity(struct idle_prof_thread *ipt)
	47	{
	48	#if defined(FIO_HAVE_CPU_AFFINITY)
	49	fio_cpuset_exit(&ipt->cpu_mask);
	50	#endif
	51	}
	52
59358c8e JA	53	static int set_cpu_affinity(struct idle_prof_thread *ipt)
	54	{
	55	#if defined(FIO_HAVE_CPU_AFFINITY)
54ed125b JA	56	if (fio_cpuset_init(&ipt->cpu_mask)) {
	57	log_err("fio: cpuset init failed\n");
	58	return -1;
	59	}
59358c8e	60
54ed125b	61	fio_cpu_set(&ipt->cpu_mask, ipt->cpu);
59358c8e	62
54ed125b	63	if (fio_setaffinity(gettid(), ipt->cpu_mask)) {
59358c8e	64	log_err("fio: fio_setaffinity failed\n");
54ed125b	65	fio_cpuset_exit(&ipt->cpu_mask);
59358c8e JA	66	return -1;
	67	}
	68
	69	return 0;
	70	#else
	71	log_err("fio: fio_setaffinity not supported\n");
	72	return -1;
	73	#endif
	74	}
	75
f2a2ce0e HL	76	static void idle_prof_thread_fn(void data)
	77	{
	78	int retval;
	79	unsigned long j, k;
	80	struct idle_prof_thread *ipt = data;
	81
	82	/* wait for all threads are spawned */
	83	pthread_mutex_lock(&ipt->init_lock);
	84
	85	/* exit if any other thread failed to start */
d5b351d4 JA	86	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
d5b351d4 JA	87	pthread_mutex_unlock(&ipt->init_lock);
f2a2ce0e	88	return NULL;
d5b351d4	89	}
f2a2ce0e	90
59358c8e	91	retval = set_cpu_affinity(ipt);
f2a2ce0e HL	92	if (retval == -1) {
	93	ipt->state = TD_EXITED;
	94	pthread_mutex_unlock(&ipt->init_lock);
	95	return NULL;
	96	}
	97
	98	ipt->cali_time = calibrate_unit(ipt->data);
	99
	100	/* delay to set IDLE class till now for better calibration accuracy */
7e09a9f1	101	#if defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	102	if ((retval = fio_set_sched_idle()))
	103	log_err("fio: fio_set_sched_idle failed\n");
	104	#else
	105	retval = -1;
	106	log_err("fio: fio_set_sched_idle not supported\n");
	107	#endif
	108	if (retval == -1) {
	109	ipt->state = TD_EXITED;
	110	pthread_mutex_unlock(&ipt->init_lock);
54ed125b	111	goto do_exit;
f2a2ce0e HL	112	}
	113
	114	ipt->state = TD_INITIALIZED;
	115
	116	/* signal the main thread that calibration is done */
	117	pthread_cond_signal(&ipt->cond);
	118	pthread_mutex_unlock(&ipt->init_lock);
	119
	120	/* wait for other calibration to finish */
	121	pthread_mutex_lock(&ipt->start_lock);
	122
	123	/* exit if other threads failed to initialize */
735ed278 JA	124	if (ipc.status == IDLE_PROF_STATUS_ABORT) {
735ed278 JA	125	pthread_mutex_unlock(&ipt->start_lock);
54ed125b	126	goto do_exit;
735ed278	127	}
f2a2ce0e HL	128
f2a2ce0e HL	129	/* exit if we are doing calibration only */
735ed278 JA	130	if (ipc.status == IDLE_PROF_STATUS_CALI_STOP) {
735ed278 JA	131	pthread_mutex_unlock(&ipt->start_lock);
54ed125b	132	goto do_exit;
735ed278	133	}
f2a2ce0e HL	134
	135	fio_gettime(&ipt->tps, NULL);
	136	ipt->state = TD_RUNNING;
	137
	138	j = 0;
	139	while (1) {
680b5abc JA	140	for (k = 0; k < page_size; k++) {
680b5abc JA	141	ipt->data[(k + j) % page_size] = k % 256;
f2a2ce0e HL	142	if (ipc.status == IDLE_PROF_STATUS_PROF_STOP) {
	143	fio_gettime(&ipt->tpe, NULL);
	144	goto idle_prof_done;
	145	}
	146	}
	147	j++;
	148	}
	149
	150	idle_prof_done:
	151
680b5abc	152	ipt->loops = j + (double) k / page_size;
f2a2ce0e HL	153	ipt->state = TD_EXITED;
	154	pthread_mutex_unlock(&ipt->start_lock);
	155
54ed125b JA	156	do_exit:
54ed125b JA	157	free_cpu_affinity(ipt);
f2a2ce0e HL	158	return NULL;
	159	}
	160
	161	/* calculate mean and standard deviation to complete an unit of work */
	162	static void calibration_stats(void)
	163	{
	164	int i;
680b5abc	165	double sum = 0.0, var = 0.0;
f2a2ce0e HL	166	struct idle_prof_thread *ipt;
	167
	168	for (i = 0; i < ipc.nr_cpus; i++) {
	169	ipt = &ipc.ipts[i];
	170	sum += ipt->cali_time;
	171	}
	172
	173	ipc.cali_mean = sum/ipc.nr_cpus;
	174
	175	for (i = 0; i < ipc.nr_cpus; i++) {
	176	ipt = &ipc.ipts[i];
	177	var += pow(ipt->cali_time-ipc.cali_mean, 2);
	178	}
	179
	180	ipc.cali_stddev = sqrt(var/(ipc.nr_cpus-1));
	181	}
	182
	183	void fio_idle_prof_init(void)
	184	{
	185	int i, ret;
	186	struct timeval tp;
	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 */
	263	log_err("fio: pthread_detatch %s\n", strerror(ret));
	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)) {
f2a2ce0e HL	285	fio_gettime(&tp, NULL);
	286	ts.tv_sec = tp.tv_sec + 1;
	287	ts.tv_nsec = tp.tv_usec * 1000;
	288	pthread_cond_timedwait(&ipt->cond, &ipt->init_lock, &ts);
	289	}
	290	pthread_mutex_unlock(&ipt->init_lock);
	291
680b5abc JA	292	/*
680b5abc JA	293	* any thread failed to initialize would abort other threads
f2a2ce0e HL	294	* later after fio_idle_prof_start.
	295	*/
	296	if (ipt->state == TD_EXITED)
	297	ipc.status = IDLE_PROF_STATUS_ABORT;
	298	}
	299
	300	if (ipc.status != IDLE_PROF_STATUS_ABORT)
	301	calibration_stats();
680b5abc	302	else
f2a2ce0e HL	303	ipc.cali_mean = ipc.cali_stddev = 0.0;
	304
	305	if (ipc.opt == IDLE_PROF_OPT_CALI)
	306	ipc.status = IDLE_PROF_STATUS_CALI_STOP;
	307	}
	308
	309	void fio_idle_prof_start(void)
	310	{
	311	int i;
	312	struct idle_prof_thread *ipt;
	313
	314	if (ipc.opt == IDLE_PROF_OPT_NONE)
	315	return;
	316
	317	/* unlock regardless abort is set or not */
	318	for (i = 0; i < ipc.nr_cpus; i++) {
	319	ipt = &ipc.ipts[i];
	320	pthread_mutex_unlock(&ipt->start_lock);
	321	}
	322	}
	323
	324	void fio_idle_prof_stop(void)
	325	{
	326	int i;
	327	uint64_t runt;
	328	struct timeval tp;
	329	struct timespec ts;
	330	struct idle_prof_thread *ipt;
	331
	332	if (ipc.opt == IDLE_PROF_OPT_NONE)
	333	return;
	334
	335	if (ipc.opt == IDLE_PROF_OPT_CALI)
	336	return;
	337
	338	ipc.status = IDLE_PROF_STATUS_PROF_STOP;
	339
	340	/* wait for all threads to exit from profiling */
	341	for (i = 0; i < ipc.nr_cpus; i++) {
	342	ipt = &ipc.ipts[i];
	343	pthread_mutex_lock(&ipt->start_lock);
680b5abc JA	344	while ((ipt->state != TD_EXITED) &&
680b5abc JA	345	(ipt->state!=TD_NOT_CREATED)) {
f2a2ce0e HL	346	fio_gettime(&tp, NULL);
	347	ts.tv_sec = tp.tv_sec + 1;
	348	ts.tv_nsec = tp.tv_usec * 1000;
	349	/* timed wait in case a signal is not received */
	350	pthread_cond_timedwait(&ipt->cond, &ipt->start_lock, &ts);
	351	}
	352	pthread_mutex_unlock(&ipt->start_lock);
	353
	354	/* calculate idleness */
	355	if (ipc.cali_mean != 0.0) {
	356	runt = utime_since(&ipt->tps, &ipt->tpe);
9da67e75 JA	357	if (runt)
	358	ipt->idleness = ipt->loops * ipc.cali_mean / runt;
	359	else
	360	ipt->idleness = 0.0;
680b5abc	361	} else
f2a2ce0e HL	362	ipt->idleness = 0.0;
	363	}
	364
680b5abc JA	365	/*
680b5abc JA	366	* memory allocations are freed via explicit fio_idle_prof_cleanup
f2a2ce0e HL	367	* after profiling stats are collected by apps.
f2a2ce0e HL	368	*/
f2a2ce0e HL	369	}
f2a2ce0e HL	370
680b5abc JA	371	/*
680b5abc JA	372	* return system idle percentage when cpu is -1;
f2a2ce0e HL	373	* return one cpu idle percentage otherwise.
	374	*/
	375	static double fio_idle_prof_cpu_stat(int cpu)
	376	{
	377	int i, nr_cpus = ipc.nr_cpus;
	378	struct idle_prof_thread *ipt;
	379	double p = 0.0;
	380
	381	if (ipc.opt == IDLE_PROF_OPT_NONE)
	382	return 0.0;
	383
	384	if ((cpu >= nr_cpus) \|\| (cpu < -1)) {
	385	log_err("fio: idle profiling invalid cpu index\n");
	386	return 0.0;
	387	}
	388
	389	if (cpu == -1) {
	390	for (i = 0; i < nr_cpus; i++) {
	391	ipt = &ipc.ipts[i];
	392	p += ipt->idleness;
	393	}
	394	p /= nr_cpus;
	395	} else {
	396	ipt = &ipc.ipts[cpu];
	397	p = ipt->idleness;
	398	}
	399
680b5abc	400	return p * 100.0;
f2a2ce0e HL	401	}
f2a2ce0e HL	402
10aa136b	403	static void fio_idle_prof_cleanup(void)
f2a2ce0e HL	404	{
	405	if (ipc.ipts) {
	406	free(ipc.ipts);
	407	ipc.ipts = NULL;
	408	}
	409
	410	if (ipc.buf) {
	411	free(ipc.buf);
	412	ipc.buf = NULL;
	413	}
	414	}
	415
	416	int fio_idle_prof_parse_opt(const char *args)
	417	{
	418	ipc.opt = IDLE_PROF_OPT_NONE; /* default */
	419
	420	if (!args) {
	421	log_err("fio: empty idle-prof option string\n");
	422	return -1;
	423	}
	424
7e09a9f1	425	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(CONFIG_SCHED_IDLE)
f2a2ce0e HL	426	if (strcmp("calibrate", args) == 0) {
	427	ipc.opt = IDLE_PROF_OPT_CALI;
	428	fio_idle_prof_init();
	429	fio_idle_prof_start();
	430	fio_idle_prof_stop();
a666cab8	431	show_idle_prof_stats(FIO_OUTPUT_NORMAL, NULL, NULL);
f2a2ce0e HL	432	return 1;
	433	} else if (strcmp("system", args) == 0) {
	434	ipc.opt = IDLE_PROF_OPT_SYSTEM;
	435	return 0;
	436	} else if (strcmp("percpu", args) == 0) {
	437	ipc.opt = IDLE_PROF_OPT_PERCPU;
	438	return 0;
	439	} else {
4e0a8fa2	440	log_err("fio: incorrect idle-prof option: %s\n", args);
f2a2ce0e HL	441	return -1;
	442	}
	443	#else
	444	log_err("fio: idle-prof not supported on this platform\n");
	445	return -1;
	446	#endif
	447	}
	448
a666cab8 JA	449	void show_idle_prof_stats(int output, struct json_object *parent,
a666cab8 JA	450	struct buf_output *out)
f2a2ce0e HL	451	{
	452	int i, nr_cpus = ipc.nr_cpus;
	453	struct json_object *tmp;
	454	char s[MAX_CPU_STR_LEN];
680b5abc	455
f2a2ce0e HL	456	if (output == FIO_OUTPUT_NORMAL) {
f2a2ce0e HL	457	if (ipc.opt > IDLE_PROF_OPT_CALI)
a666cab8	458	log_buf(out, "\nCPU idleness:\n");
f2a2ce0e	459	else if (ipc.opt == IDLE_PROF_OPT_CALI)
a666cab8	460	log_buf(out, "CPU idleness:\n");
f2a2ce0e HL	461
f2a2ce0e HL	462	if (ipc.opt >= IDLE_PROF_OPT_SYSTEM)
a666cab8	463	log_buf(out, " system: %3.2f%%\n", fio_idle_prof_cpu_stat(-1));
f2a2ce0e HL	464
f2a2ce0e HL	465	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
a666cab8	466	log_buf(out, " percpu: %3.2f%%", fio_idle_prof_cpu_stat(0));
680b5abc	467	for (i = 1; i < nr_cpus; i++)
a666cab8 JA	468	log_buf(out, ", %3.2f%%", fio_idle_prof_cpu_stat(i));
a666cab8 JA	469	log_buf(out, "\n");
f2a2ce0e HL	470	}
	471
	472	if (ipc.opt >= IDLE_PROF_OPT_CALI) {
a666cab8 JA	473	log_buf(out, " unit work: mean=%3.2fus,", ipc.cali_mean);
a666cab8 JA	474	log_buf(out, " stddev=%3.2f\n", ipc.cali_stddev);
f2a2ce0e HL	475	}
	476
	477	/* dynamic mem allocations can now be freed */
	478	if (ipc.opt != IDLE_PROF_OPT_NONE)
	479	fio_idle_prof_cleanup();
	480
	481	return;
	482	}
680b5abc	483
236d24df	484	if ((ipc.opt != IDLE_PROF_OPT_NONE) && (output & FIO_OUTPUT_JSON)) {
f2a2ce0e HL	485	if (!parent)
	486	return;
	487
	488	tmp = json_create_object();
	489	if (!tmp)
	490	return;
	491
	492	json_object_add_value_object(parent, "cpu_idleness", tmp);
	493	json_object_add_value_float(tmp, "system", fio_idle_prof_cpu_stat(-1));
	494
	495	if (ipc.opt == IDLE_PROF_OPT_PERCPU) {
680b5abc	496	for (i = 0; i < nr_cpus; i++) {
f2a2ce0e HL	497	snprintf(s, MAX_CPU_STR_LEN, "cpu-%d", i);
	498	json_object_add_value_float(tmp, s, fio_idle_prof_cpu_stat(i));
	499	}
	500	}
	501
	502	json_object_add_value_float(tmp, "unit_mean", ipc.cali_mean);
	503	json_object_add_value_float(tmp, "unit_stddev", ipc.cali_stddev);
	504
	505	fio_idle_prof_cleanup();
f2a2ce0e HL	506	}
f2a2ce0e HL	507	}