[fio.git] / gettime.c

/*
 * Clock functions
 */

#include <unistd.h>
#include <math.h>
#include <sys/time.h>
#include <time.h>

#include "fio.h"
#include "smalloc.h"

#include "hash.h"
#include "os/os.h"

#ifdef ARCH_HAVE_CPU_CLOCK
static unsigned long cycles_per_usec;
static unsigned long inv_cycles_per_usec;
#endif
int tsc_reliable = 0;

struct tv_valid {
	struct timeval last_tv;
	uint64_t last_cycles;
	int last_tv_valid;
};
#ifdef CONFIG_TLS_THREAD
static __thread struct tv_valid static_tv_valid;
#else
static pthread_key_t tv_tls_key;
#endif

enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
int fio_clock_source_set = 0;
enum fio_cs fio_clock_source_inited = CS_INVAL;

#ifdef FIO_DEBUG_TIME

#define HASH_BITS	8
#define HASH_SIZE	(1 << HASH_BITS)

static struct flist_head hash[HASH_SIZE];
static int gtod_inited;

struct gtod_log {
	struct flist_head list;
	void *caller;
	unsigned long calls;
};

static struct gtod_log *find_hash(void *caller)
{
	unsigned long h = hash_ptr(caller, HASH_BITS);
	struct flist_head *entry;

	flist_for_each(entry, &hash[h]) {
		struct gtod_log *log = flist_entry(entry, struct gtod_log,
									list);

		if (log->caller == caller)
			return log;
	}

	return NULL;
}

static struct gtod_log *find_log(void *caller)
{
	struct gtod_log *log = find_hash(caller);

	if (!log) {
		unsigned long h;

		log = malloc(sizeof(*log));
		INIT_FLIST_HEAD(&log->list);
		log->caller = caller;
		log->calls = 0;

		h = hash_ptr(caller, HASH_BITS);
		flist_add_tail(&log->list, &hash[h]);
	}

	return log;
}

static void gtod_log_caller(void *caller)
{
	if (gtod_inited) {
		struct gtod_log *log = find_log(caller);

		log->calls++;
	}
}

static void fio_exit fio_dump_gtod(void)
{
	unsigned long total_calls = 0;
	int i;

	for (i = 0; i < HASH_SIZE; i++) {
		struct flist_head *entry;
		struct gtod_log *log;

		flist_for_each(entry, &hash[i]) {
			log = flist_entry(entry, struct gtod_log, list);

			printf("function %p, calls %lu\n", log->caller,
								log->calls);
			total_calls += log->calls;
		}
	}

	printf("Total %lu gettimeofday\n", total_calls);
}

static void fio_init gtod_init(void)
{
	int i;

	for (i = 0; i < HASH_SIZE; i++)
		INIT_FLIST_HEAD(&hash[i]);

	gtod_inited = 1;
}

#endif /* FIO_DEBUG_TIME */

#ifdef CONFIG_CLOCK_GETTIME
static int fill_clock_gettime(struct timespec *ts)
{
#ifdef CONFIG_CLOCK_MONOTONIC
	return clock_gettime(CLOCK_MONOTONIC, ts);
#else
	return clock_gettime(CLOCK_REALTIME, ts);
#endif
}
#endif

static void *__fio_gettime(struct timeval *tp)
{
	struct tv_valid *tv;

#ifdef CONFIG_TLS_THREAD
	tv = &static_tv_valid;
#else
	tv = pthread_getspecific(tv_tls_key);
#endif

	switch (fio_clock_source) {
#ifdef CONFIG_GETTIMEOFDAY
	case CS_GTOD:
		gettimeofday(tp, NULL);
		break;
#endif
#ifdef CONFIG_CLOCK_GETTIME
	case CS_CGETTIME: {
		struct timespec ts;

		if (fill_clock_gettime(&ts) < 0) {
			log_err("fio: clock_gettime fails\n");
			assert(0);
		}

		tp->tv_sec = ts.tv_sec;
		tp->tv_usec = ts.tv_nsec / 1000;
		break;
		}
#endif
#ifdef ARCH_HAVE_CPU_CLOCK
	case CS_CPUCLOCK: {
		uint64_t usecs, t;

		t = get_cpu_clock();
		if (tv && t < tv->last_cycles) {
			dprint(FD_TIME, "CPU clock going back in time\n");
			t = tv->last_cycles;
		} else if (tv)
			tv->last_cycles = t;

		usecs = (t * inv_cycles_per_usec) / 16777216UL;
		tp->tv_sec = usecs / 1000000;
		tp->tv_usec = usecs % 1000000;
		break;
		}
#endif
	default:
		log_err("fio: invalid clock source %d\n", fio_clock_source);
		break;
	}

	return tv;
}

#ifdef FIO_DEBUG_TIME
void fio_gettime(struct timeval *tp, void *caller)
#else
void fio_gettime(struct timeval *tp, void fio_unused *caller)
#endif
{
	struct tv_valid *tv;

#ifdef FIO_DEBUG_TIME
	if (!caller)
		caller = __builtin_return_address(0);

	gtod_log_caller(caller);
#endif
	if (fio_tv) {
		memcpy(tp, fio_tv, sizeof(*tp));
		return;
	}

	tv = __fio_gettime(tp);

	/*
	 * If Linux is using the tsc clock on non-synced processors,
	 * sometimes time can appear to drift backwards. Fix that up.
	 */
	if (tv) {
		if (tv->last_tv_valid) {
			if (tp->tv_sec < tv->last_tv.tv_sec)
				tp->tv_sec = tv->last_tv.tv_sec;
			else if (tv->last_tv.tv_sec == tp->tv_sec &&
				 tp->tv_usec < tv->last_tv.tv_usec)
				tp->tv_usec = tv->last_tv.tv_usec;
		}
		tv->last_tv_valid = 1;
		memcpy(&tv->last_tv, tp, sizeof(*tp));
	}
}

#ifdef ARCH_HAVE_CPU_CLOCK
static unsigned long get_cycles_per_usec(void)
{
	struct timeval s, e;
	uint64_t c_s, c_e;
	enum fio_cs old_cs = fio_clock_source;

#ifdef CONFIG_CLOCK_GETTIME
	fio_clock_source = CS_CGETTIME;
#else
	fio_clock_source = CS_GTOD;
#endif
	__fio_gettime(&s);

	c_s = get_cpu_clock();
	do {
		uint64_t elapsed;

		__fio_gettime(&e);

		elapsed = utime_since(&s, &e);
		if (elapsed >= 1280) {
			c_e = get_cpu_clock();
			break;
		}
	} while (1);

	fio_clock_source = old_cs;
	return (c_e - c_s + 127) >> 7;
}

#define NR_TIME_ITERS	50

static int calibrate_cpu_clock(void)
{
	double delta, mean, S;
	uint64_t avg, cycles[NR_TIME_ITERS];
	int i, samples;

	cycles[0] = get_cycles_per_usec();
	S = delta = mean = 0.0;
	for (i = 0; i < NR_TIME_ITERS; i++) {
		cycles[i] = get_cycles_per_usec();
		delta = cycles[i] - mean;
		if (delta) {
			mean += delta / (i + 1.0);
			S += delta * (cycles[i] - mean);
		}
	}

	/*
	 * The most common platform clock breakage is returning zero
	 * indefinitely. Check for that and return failure.
	 */
	if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
		return 1;

	S = sqrt(S / (NR_TIME_ITERS - 1.0));

	samples = avg = 0;
	for (i = 0; i < NR_TIME_ITERS; i++) {
		double this = cycles[i];

		if ((fmax(this, mean) - fmin(this, mean)) > S)
			continue;
		samples++;
		avg += this;
	}

	S /= (double) NR_TIME_ITERS;
	mean /= 10.0;

	for (i = 0; i < NR_TIME_ITERS; i++)
		dprint(FD_TIME, "cycles[%d]=%llu\n", i,
					(unsigned long long) cycles[i] / 10);

	avg /= samples;
	avg = (avg + 5) / 10;
	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);

	cycles_per_usec = avg;
	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
	return 0;
}
#else
static int calibrate_cpu_clock(void)
{
	return 1;
}
#endif

#ifndef CONFIG_TLS_THREAD
void fio_local_clock_init(int is_thread)
{
	struct tv_valid *t;

	t = calloc(sizeof(*t), 1);
	if (pthread_setspecific(tv_tls_key, t))
		log_err("fio: can't set TLS key\n");
}

static void kill_tv_tls_key(void *data)
{
	free(data);
}
#else
void fio_local_clock_init(int is_thread)
{
}
#endif

void fio_clock_init(void)
{
	if (fio_clock_source == fio_clock_source_inited)
		return;

#ifndef CONFIG_TLS_THREAD
	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
		log_err("fio: can't create TLS key\n");
#endif

	fio_clock_source_inited = fio_clock_source;

	if (calibrate_cpu_clock())
		tsc_reliable = 0;

	/*
	 * If the arch sets tsc_reliable != 0, then it must be good enough
	 * to use as THE clock source. For x86 CPUs, this means the TSC
	 * runs at a constant rate and is synced across CPU cores.
	 */
	if (tsc_reliable) {
		if (!fio_clock_source_set)
			fio_clock_source = CS_CPUCLOCK;
	} else if (fio_clock_source == CS_CPUCLOCK)
		log_info("fio: clocksource=cpu may not be reliable\n");
}

uint64_t utime_since(struct timeval *s, struct timeval *e)
{
	long sec, usec;
	uint64_t ret;

	sec = e->tv_sec - s->tv_sec;
	usec = e->tv_usec - s->tv_usec;
	if (sec > 0 && usec < 0) {
		sec--;
		usec += 1000000;
	}

	/*
	 * time warp bug on some kernels?
	 */
	if (sec < 0 || (sec == 0 && usec < 0))
		return 0;

	ret = sec * 1000000ULL + usec;

	return ret;
}

uint64_t utime_since_now(struct timeval *s)
{
	struct timeval t;

	fio_gettime(&t, NULL);
	return utime_since(s, &t);
}

uint64_t mtime_since(struct timeval *s, struct timeval *e)
{
	long sec, usec, ret;

	sec = e->tv_sec - s->tv_sec;
	usec = e->tv_usec - s->tv_usec;
	if (sec > 0 && usec < 0) {
		sec--;
		usec += 1000000;
	}

	if (sec < 0 || (sec == 0 && usec < 0))
		return 0;

	sec *= 1000UL;
	usec /= 1000UL;
	ret = sec + usec;

	return ret;
}

uint64_t mtime_since_now(struct timeval *s)
{
	struct timeval t;
	void *p = __builtin_return_address(0);

	fio_gettime(&t, p);
	return mtime_since(s, &t);
}

uint64_t time_since_now(struct timeval *s)
{
	return mtime_since_now(s) / 1000;
}

#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
    defined(CONFIG_SFAA)

#define CLOCK_ENTRIES	100000

struct clock_entry {
	uint32_t seq;
	uint32_t cpu;
	uint64_t tsc;
};

struct clock_thread {
	pthread_t thread;
	int cpu;
	pthread_mutex_t lock;
	pthread_mutex_t started;
	uint32_t *seq;
	struct clock_entry *entries;
};

static inline uint32_t atomic32_inc_return(uint32_t *seq)
{
	return 1 + __sync_fetch_and_add(seq, 1);
}

static void *clock_thread_fn(void *data)
{
	struct clock_thread *t = data;
	struct clock_entry *c;
	os_cpu_mask_t cpu_mask;
	uint32_t last_seq;
	int i;

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

	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
		log_err("clock setaffinity failed\n");
		return (void *) 1;
	}

	pthread_mutex_lock(&t->lock);
	pthread_mutex_unlock(&t->started);

	last_seq = 0;
	c = &t->entries[0];
	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
		uint32_t seq;
		uint64_t tsc;

		c->cpu = t->cpu;
		do {
			seq = atomic32_inc_return(t->seq);
			if (seq < last_seq)
				break;
			tsc = get_cpu_clock();
		} while (seq != *t->seq);

		c->seq = seq;
		c->tsc = tsc;
	}

	log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
		(unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);

	/*
	 * The most common platform clock breakage is returning zero
	 * indefinitely. Check for that and return failure.
	 */
	if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
		return (void *) 1;

	return NULL;
}

static int clock_cmp(const void *p1, const void *p2)
{
	const struct clock_entry *c1 = p1;
	const struct clock_entry *c2 = p2;

	if (c1->seq == c2->seq)
		log_err("cs: bug in atomic sequence!\n");

	return c1->seq - c2->seq;
}

int fio_monotonic_clocktest(void)
{
	struct clock_thread *threads;
	unsigned int nr_cpus = cpus_online();
	struct clock_entry *entries;
	unsigned long tentries, failed;
	struct clock_entry *prev, *this;
	uint32_t seq = 0;
	int i;

	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");

	fio_debug |= 1U << FD_TIME;
	calibrate_cpu_clock();
	fio_debug &= ~(1U << FD_TIME);

	threads = malloc(nr_cpus * sizeof(struct clock_thread));
	tentries = CLOCK_ENTRIES * nr_cpus;
	entries = malloc(tentries * sizeof(struct clock_entry));

	log_info("cs: Testing %u CPUs\n", nr_cpus);

	for (i = 0; i < nr_cpus; i++) {
		struct clock_thread *t = &threads[i];

		t->cpu = i;
		t->seq = &seq;
		t->entries = &entries[i * CLOCK_ENTRIES];
		pthread_mutex_init(&t->lock, NULL);
		pthread_mutex_init(&t->started, NULL);
		pthread_mutex_lock(&t->lock);
		pthread_create(&t->thread, NULL, clock_thread_fn, t);
	}

	for (i = 0; i < nr_cpus; i++) {
		struct clock_thread *t = &threads[i];

		pthread_mutex_lock(&t->started);
	}

	for (i = 0; i < nr_cpus; i++) {
		struct clock_thread *t = &threads[i];

		pthread_mutex_unlock(&t->lock);
	}

	for (failed = i = 0; i < nr_cpus; i++) {
		struct clock_thread *t = &threads[i];
		void *ret;

		pthread_join(t->thread, &ret);
		if (ret)
			failed++;
	}
	free(threads);

	if (failed) {
		log_err("Clocksource test: %u threads failed\n", failed);
		goto err;
	}

	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);

	for (failed = i = 0; i < tentries; i++) {
		this = &entries[i];

		if (!i) {
			prev = this;
			continue;
		}

		if (prev->tsc > this->tsc) {
			uint64_t diff = prev->tsc - this->tsc;

			log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
			log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
			log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
			failed++;
		}

		prev = this;
	}

	if (failed)
		log_info("cs: Failed: %lu\n", failed);
	else
		log_info("cs: Pass!\n");

err:
	free(entries);
	return !!failed;
}

#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */

int fio_monotonic_clocktest(void)
{
	log_info("cs: current platform does not support CPU clocks\n");
	return 0;
}

#endif
Commit	Line	Data
	1	/*
	2	* Clock functions
	3	*/
	4
	5	#include <unistd.h>
	6	#include <math.h>
	7	#include <sys/time.h>
	8	#include <time.h>
	9
	10	#include "fio.h"
	11	#include "smalloc.h"
	12
	13	#include "hash.h"
	14	#include "os/os.h"
	15
	16	#ifdef ARCH_HAVE_CPU_CLOCK
	17	static unsigned long cycles_per_usec;
	18	static unsigned long inv_cycles_per_usec;
	19	#endif
	20	int tsc_reliable = 0;
	21
	22	struct tv_valid {
	23	struct timeval last_tv;
	24	uint64_t last_cycles;
	25	int last_tv_valid;
	26	};
	27	#ifdef CONFIG_TLS_THREAD
	28	static __thread struct tv_valid static_tv_valid;
	29	#else
	30	static pthread_key_t tv_tls_key;
	31	#endif
	32
	33	enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
	34	int fio_clock_source_set = 0;
	35	enum fio_cs fio_clock_source_inited = CS_INVAL;
	36
	37	#ifdef FIO_DEBUG_TIME
	38
	39	#define HASH_BITS 8
	40	#define HASH_SIZE (1 << HASH_BITS)
	41
	42	static struct flist_head hash[HASH_SIZE];
	43	static int gtod_inited;
	44
	45	struct gtod_log {
	46	struct flist_head list;
	47	void *caller;
	48	unsigned long calls;
	49	};
	50
	51	static struct gtod_log find_hash(void caller)
	52	{
	53	unsigned long h = hash_ptr(caller, HASH_BITS);
	54	struct flist_head *entry;
	55
	56	flist_for_each(entry, &hash[h]) {
	57	struct gtod_log *log = flist_entry(entry, struct gtod_log,
	58	list);
	59
	60	if (log->caller == caller)
	61	return log;
	62	}
	63
	64	return NULL;
	65	}
	66
	67	static struct gtod_log find_log(void caller)
	68	{
	69	struct gtod_log *log = find_hash(caller);
	70
	71	if (!log) {
	72	unsigned long h;
	73
	74	log = malloc(sizeof(*log));
	75	INIT_FLIST_HEAD(&log->list);
	76	log->caller = caller;
	77	log->calls = 0;
	78
	79	h = hash_ptr(caller, HASH_BITS);
	80	flist_add_tail(&log->list, &hash[h]);
	81	}
	82
	83	return log;
	84	}
	85
	86	static void gtod_log_caller(void *caller)
	87	{
	88	if (gtod_inited) {
	89	struct gtod_log *log = find_log(caller);
	90
	91	log->calls++;
	92	}
	93	}
	94
	95	static void fio_exit fio_dump_gtod(void)
	96	{
	97	unsigned long total_calls = 0;
	98	int i;
	99
	100	for (i = 0; i < HASH_SIZE; i++) {
	101	struct flist_head *entry;
	102	struct gtod_log *log;
	103
	104	flist_for_each(entry, &hash[i]) {
	105	log = flist_entry(entry, struct gtod_log, list);
	106
	107	printf("function %p, calls %lu\n", log->caller,
	108	log->calls);
	109	total_calls += log->calls;
	110	}
	111	}
	112
	113	printf("Total %lu gettimeofday\n", total_calls);
	114	}
	115
	116	static void fio_init gtod_init(void)
	117	{
	118	int i;
	119
	120	for (i = 0; i < HASH_SIZE; i++)
	121	INIT_FLIST_HEAD(&hash[i]);
	122
	123	gtod_inited = 1;
	124	}
	125
	126	#endif /* FIO_DEBUG_TIME */
	127
	128	#ifdef CONFIG_CLOCK_GETTIME
	129	static int fill_clock_gettime(struct timespec *ts)
	130	{
	131	#ifdef CONFIG_CLOCK_MONOTONIC
	132	return clock_gettime(CLOCK_MONOTONIC, ts);
	133	#else
	134	return clock_gettime(CLOCK_REALTIME, ts);
	135	#endif
	136	}
	137	#endif
	138
	139	static void __fio_gettime(struct timeval tp)
	140	{
	141	struct tv_valid *tv;
	142
	143	#ifdef CONFIG_TLS_THREAD
	144	tv = &static_tv_valid;
	145	#else
	146	tv = pthread_getspecific(tv_tls_key);
	147	#endif
	148
	149	switch (fio_clock_source) {
	150	#ifdef CONFIG_GETTIMEOFDAY
	151	case CS_GTOD:
	152	gettimeofday(tp, NULL);
	153	break;
	154	#endif
	155	#ifdef CONFIG_CLOCK_GETTIME
	156	case CS_CGETTIME: {
	157	struct timespec ts;
	158
	159	if (fill_clock_gettime(&ts) < 0) {
	160	log_err("fio: clock_gettime fails\n");
	161	assert(0);
	162	}
	163
	164	tp->tv_sec = ts.tv_sec;
	165	tp->tv_usec = ts.tv_nsec / 1000;
	166	break;
	167	}
	168	#endif
	169	#ifdef ARCH_HAVE_CPU_CLOCK
	170	case CS_CPUCLOCK: {
	171	uint64_t usecs, t;
	172
	173	t = get_cpu_clock();
	174	if (tv && t < tv->last_cycles) {
	175	dprint(FD_TIME, "CPU clock going back in time\n");
	176	t = tv->last_cycles;
	177	} else if (tv)
	178	tv->last_cycles = t;
	179
	180	usecs = (t * inv_cycles_per_usec) / 16777216UL;
	181	tp->tv_sec = usecs / 1000000;
	182	tp->tv_usec = usecs % 1000000;
	183	break;
	184	}
	185	#endif
	186	default:
	187	log_err("fio: invalid clock source %d\n", fio_clock_source);
	188	break;
	189	}
	190
	191	return tv;
	192	}
	193
	194	#ifdef FIO_DEBUG_TIME
	195	void fio_gettime(struct timeval tp, void caller)
	196	#else
	197	void fio_gettime(struct timeval tp, void fio_unused caller)
	198	#endif
	199	{
	200	struct tv_valid *tv;
	201
	202	#ifdef FIO_DEBUG_TIME
	203	if (!caller)
	204	caller = __builtin_return_address(0);
	205
	206	gtod_log_caller(caller);
	207	#endif
	208	if (fio_tv) {
	209	memcpy(tp, fio_tv, sizeof(*tp));
	210	return;
	211	}
	212
	213	tv = __fio_gettime(tp);
	214
	215	/*
	216	* If Linux is using the tsc clock on non-synced processors,
	217	* sometimes time can appear to drift backwards. Fix that up.
	218	*/
	219	if (tv) {
	220	if (tv->last_tv_valid) {
	221	if (tp->tv_sec < tv->last_tv.tv_sec)
	222	tp->tv_sec = tv->last_tv.tv_sec;
	223	else if (tv->last_tv.tv_sec == tp->tv_sec &&
	224	tp->tv_usec < tv->last_tv.tv_usec)
	225	tp->tv_usec = tv->last_tv.tv_usec;
	226	}
	227	tv->last_tv_valid = 1;
	228	memcpy(&tv->last_tv, tp, sizeof(*tp));
	229	}
	230	}
	231
	232	#ifdef ARCH_HAVE_CPU_CLOCK
	233	static unsigned long get_cycles_per_usec(void)
	234	{
	235	struct timeval s, e;
	236	uint64_t c_s, c_e;
	237	enum fio_cs old_cs = fio_clock_source;
	238
	239	#ifdef CONFIG_CLOCK_GETTIME
	240	fio_clock_source = CS_CGETTIME;
	241	#else
	242	fio_clock_source = CS_GTOD;
	243	#endif
	244	__fio_gettime(&s);
	245
	246	c_s = get_cpu_clock();
	247	do {
	248	uint64_t elapsed;
	249
	250	__fio_gettime(&e);
	251
	252	elapsed = utime_since(&s, &e);
	253	if (elapsed >= 1280) {
	254	c_e = get_cpu_clock();
	255	break;
	256	}
	257	} while (1);
	258
	259	fio_clock_source = old_cs;
	260	return (c_e - c_s + 127) >> 7;
	261	}
	262
	263	#define NR_TIME_ITERS 50
	264
	265	static int calibrate_cpu_clock(void)
	266	{
	267	double delta, mean, S;
	268	uint64_t avg, cycles[NR_TIME_ITERS];
	269	int i, samples;
	270
	271	cycles[0] = get_cycles_per_usec();
	272	S = delta = mean = 0.0;
	273	for (i = 0; i < NR_TIME_ITERS; i++) {
	274	cycles[i] = get_cycles_per_usec();
	275	delta = cycles[i] - mean;
	276	if (delta) {
	277	mean += delta / (i + 1.0);
	278	S += delta * (cycles[i] - mean);
	279	}
	280	}
	281
	282	/*
	283	* The most common platform clock breakage is returning zero
	284	* indefinitely. Check for that and return failure.
	285	*/
	286	if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
	287	return 1;
	288
	289	S = sqrt(S / (NR_TIME_ITERS - 1.0));
	290
	291	samples = avg = 0;
	292	for (i = 0; i < NR_TIME_ITERS; i++) {
	293	double this = cycles[i];
	294
	295	if ((fmax(this, mean) - fmin(this, mean)) > S)
	296	continue;
	297	samples++;
	298	avg += this;
	299	}
	300
	301	S /= (double) NR_TIME_ITERS;
	302	mean /= 10.0;
	303
	304	for (i = 0; i < NR_TIME_ITERS; i++)
	305	dprint(FD_TIME, "cycles[%d]=%llu\n", i,
	306	(unsigned long long) cycles[i] / 10);
	307
	308	avg /= samples;
	309	avg = (avg + 5) / 10;
	310	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
	311	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
	312
	313	cycles_per_usec = avg;
	314	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
	315	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
	316	return 0;
	317	}
	318	#else
	319	static int calibrate_cpu_clock(void)
	320	{
	321	return 1;
	322	}
	323	#endif
	324
	325	#ifndef CONFIG_TLS_THREAD
	326	void fio_local_clock_init(int is_thread)
	327	{
	328	struct tv_valid *t;
	329
	330	t = calloc(sizeof(*t), 1);
	331	if (pthread_setspecific(tv_tls_key, t))
	332	log_err("fio: can't set TLS key\n");
	333	}
	334
	335	static void kill_tv_tls_key(void *data)
	336	{
	337	free(data);
	338	}
	339	#else
	340	void fio_local_clock_init(int is_thread)
	341	{
	342	}
	343	#endif
	344
	345	void fio_clock_init(void)
	346	{
	347	if (fio_clock_source == fio_clock_source_inited)
	348	return;
	349
	350	#ifndef CONFIG_TLS_THREAD
	351	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
	352	log_err("fio: can't create TLS key\n");
	353	#endif
	354
	355	fio_clock_source_inited = fio_clock_source;
	356
	357	if (calibrate_cpu_clock())
	358	tsc_reliable = 0;
	359
	360	/*
	361	* If the arch sets tsc_reliable != 0, then it must be good enough
	362	* to use as THE clock source. For x86 CPUs, this means the TSC
	363	* runs at a constant rate and is synced across CPU cores.
	364	*/
	365	if (tsc_reliable) {
	366	if (!fio_clock_source_set)
	367	fio_clock_source = CS_CPUCLOCK;
	368	} else if (fio_clock_source == CS_CPUCLOCK)
	369	log_info("fio: clocksource=cpu may not be reliable\n");
	370	}
	371
	372	uint64_t utime_since(struct timeval s, struct timeval e)
	373	{
	374	long sec, usec;
	375	uint64_t ret;
	376
	377	sec = e->tv_sec - s->tv_sec;
	378	usec = e->tv_usec - s->tv_usec;
	379	if (sec > 0 && usec < 0) {
	380	sec--;
	381	usec += 1000000;
	382	}
	383
	384	/*
	385	* time warp bug on some kernels?
	386	*/
	387	if (sec < 0 \|\| (sec == 0 && usec < 0))
	388	return 0;
	389
	390	ret = sec * 1000000ULL + usec;
	391
	392	return ret;
	393	}
	394
	395	uint64_t utime_since_now(struct timeval *s)
	396	{
	397	struct timeval t;
	398
	399	fio_gettime(&t, NULL);
	400	return utime_since(s, &t);
	401	}
	402
	403	uint64_t mtime_since(struct timeval s, struct timeval e)
	404	{
	405	long sec, usec, ret;
	406
	407	sec = e->tv_sec - s->tv_sec;
	408	usec = e->tv_usec - s->tv_usec;
	409	if (sec > 0 && usec < 0) {
	410	sec--;
	411	usec += 1000000;
	412	}
	413
	414	if (sec < 0 \|\| (sec == 0 && usec < 0))
	415	return 0;
	416
	417	sec *= 1000UL;
	418	usec /= 1000UL;
	419	ret = sec + usec;
	420
	421	return ret;
	422	}
	423
	424	uint64_t mtime_since_now(struct timeval *s)
	425	{
	426	struct timeval t;
	427	void *p = __builtin_return_address(0);
	428
	429	fio_gettime(&t, p);
	430	return mtime_since(s, &t);
	431	}
	432
	433	uint64_t time_since_now(struct timeval *s)
	434	{
	435	return mtime_since_now(s) / 1000;
	436	}
	437
	438	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
	439	defined(CONFIG_SFAA)
	440
	441	#define CLOCK_ENTRIES 100000
	442
	443	struct clock_entry {
	444	uint32_t seq;
	445	uint32_t cpu;
	446	uint64_t tsc;
	447	};
	448
	449	struct clock_thread {
	450	pthread_t thread;
	451	int cpu;
	452	pthread_mutex_t lock;
	453	pthread_mutex_t started;
	454	uint32_t *seq;
	455	struct clock_entry *entries;
	456	};
	457
	458	static inline uint32_t atomic32_inc_return(uint32_t *seq)
	459	{
	460	return 1 + __sync_fetch_and_add(seq, 1);
	461	}
	462
	463	static void clock_thread_fn(void data)
	464	{
	465	struct clock_thread *t = data;
	466	struct clock_entry *c;
	467	os_cpu_mask_t cpu_mask;
	468	uint32_t last_seq;
	469	int i;
	470
	471	memset(&cpu_mask, 0, sizeof(cpu_mask));
	472	fio_cpu_set(&cpu_mask, t->cpu);
	473
	474	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
	475	log_err("clock setaffinity failed\n");
	476	return (void *) 1;
	477	}
	478
	479	pthread_mutex_lock(&t->lock);
	480	pthread_mutex_unlock(&t->started);
	481
	482	last_seq = 0;
	483	c = &t->entries[0];
	484	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
	485	uint32_t seq;
	486	uint64_t tsc;
	487
	488	c->cpu = t->cpu;
	489	do {
	490	seq = atomic32_inc_return(t->seq);
	491	if (seq < last_seq)
	492	break;
	493	tsc = get_cpu_clock();
	494	} while (seq != *t->seq);
	495
	496	c->seq = seq;
	497	c->tsc = tsc;
	498	}
	499
	500	log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
	501	(unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
	502
	503	/*
	504	* The most common platform clock breakage is returning zero
	505	* indefinitely. Check for that and return failure.
	506	*/
	507	if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
	508	return (void *) 1;
	509
	510	return NULL;
	511	}
	512
	513	static int clock_cmp(const void p1, const void p2)
	514	{
	515	const struct clock_entry *c1 = p1;
	516	const struct clock_entry *c2 = p2;
	517
	518	if (c1->seq == c2->seq)
	519	log_err("cs: bug in atomic sequence!\n");
	520
	521	return c1->seq - c2->seq;
	522	}
	523
	524	int fio_monotonic_clocktest(void)
	525	{
	526	struct clock_thread *threads;
	527	unsigned int nr_cpus = cpus_online();
	528	struct clock_entry *entries;
	529	unsigned long tentries, failed;
	530	struct clock_entry prev, this;
	531	uint32_t seq = 0;
	532	int i;
	533
	534	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
	535
	536	fio_debug \|= 1U << FD_TIME;
	537	calibrate_cpu_clock();
	538	fio_debug &= ~(1U << FD_TIME);
	539
	540	threads = malloc(nr_cpus * sizeof(struct clock_thread));
	541	tentries = CLOCK_ENTRIES * nr_cpus;
	542	entries = malloc(tentries * sizeof(struct clock_entry));
	543
	544	log_info("cs: Testing %u CPUs\n", nr_cpus);
	545
	546	for (i = 0; i < nr_cpus; i++) {
	547	struct clock_thread *t = &threads[i];
	548
	549	t->cpu = i;
	550	t->seq = &seq;
	551	t->entries = &entries[i * CLOCK_ENTRIES];
	552	pthread_mutex_init(&t->lock, NULL);
	553	pthread_mutex_init(&t->started, NULL);
	554	pthread_mutex_lock(&t->lock);
	555	pthread_create(&t->thread, NULL, clock_thread_fn, t);
	556	}
	557
	558	for (i = 0; i < nr_cpus; i++) {
	559	struct clock_thread *t = &threads[i];
	560
	561	pthread_mutex_lock(&t->started);
	562	}
	563
	564	for (i = 0; i < nr_cpus; i++) {
	565	struct clock_thread *t = &threads[i];
	566
	567	pthread_mutex_unlock(&t->lock);
	568	}
	569
	570	for (failed = i = 0; i < nr_cpus; i++) {
	571	struct clock_thread *t = &threads[i];
	572	void *ret;
	573
	574	pthread_join(t->thread, &ret);
	575	if (ret)
	576	failed++;
	577	}
	578	free(threads);
	579
	580	if (failed) {
	581	log_err("Clocksource test: %u threads failed\n", failed);
	582	goto err;
	583	}
	584
	585	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	586
	587	for (failed = i = 0; i < tentries; i++) {
	588	this = &entries[i];
	589
	590	if (!i) {
	591	prev = this;
	592	continue;
	593	}
	594
	595	if (prev->tsc > this->tsc) {
	596	uint64_t diff = prev->tsc - this->tsc;
	597
	598	log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
	599	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
	600	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
	601	failed++;
	602	}
	603
	604	prev = this;
	605	}
	606
	607	if (failed)
	608	log_info("cs: Failed: %lu\n", failed);
	609	else
	610	log_info("cs: Pass!\n");
	611
	612	err:
	613	free(entries);
	614	return !!failed;
	615	}
	616
	617	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
	618
	619	int fio_monotonic_clocktest(void)
	620	{
	621	log_info("cs: current platform does not support CPU clocks\n");
	622	return 0;
	623	}
	624
	625	#endif