[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;
int tsc_reliable = 0;
#endif

struct tv_valid {
	struct timeval last_tv;
	int last_tv_valid;
	unsigned long last_cycles;
};
static pthread_key_t tv_tls_key;

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 */

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

#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 = pthread_getspecific(tv_tls_key);

	switch (fio_clock_source) {
	case CS_GTOD:
		gettimeofday(tp, NULL);
		break;
	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;
		}
#ifdef ARCH_HAVE_CPU_CLOCK
	case CS_CPUCLOCK: {
		unsigned long long 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;
	}

	/*
	 * 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 timespec ts;
	struct timeval s, e;
	unsigned long long c_s, c_e;

	fill_clock_gettime(&ts);
	s.tv_sec = ts.tv_sec;
	s.tv_usec = ts.tv_nsec / 1000;

	c_s = get_cpu_clock();
	do {
		unsigned long long elapsed;

		fill_clock_gettime(&ts);
		e.tv_sec = ts.tv_sec;
		e.tv_usec = ts.tv_nsec / 1000;

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

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

#define NR_TIME_ITERS	50

static void calibrate_cpu_clock(void)
{
	double delta, mean, S;
	unsigned long 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);
		}
	}

	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]=%lu\n", i, cycles[i] / 10);

	avg /= samples;
	avg = (avg + 9) / 10;
	dprint(FD_TIME, "avg: %lu\n", 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);
}
#else
static void calibrate_cpu_clock(void)
{
}
#endif

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);
}

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

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

	fio_clock_source_inited = fio_clock_source;
	calibrate_cpu_clock();

	/*
	 * 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");
}

unsigned long long utime_since(struct timeval *s, struct timeval *e)
{
	long sec, usec;
	unsigned long long 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;
}

unsigned long long utime_since_now(struct timeval *s)
{
	struct timeval t;

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

unsigned long 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;
}

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

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

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

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

#define CLOCK_ENTRIES	100000

struct clock_entry {
	unsigned long seq;
	unsigned long tsc;
	unsigned long cpu;
};

struct clock_thread {
	pthread_t thread;
	int cpu;
	pthread_mutex_t lock;
	pthread_mutex_t started;
	uint64_t *seq;
	struct clock_entry *entries;
};

static inline uint64_t atomic64_inc_return(uint64_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;
	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);

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

		c->cpu = t->cpu;
		do {
			seq = atomic64_inc_return(t->seq);
			tsc = get_cpu_clock();
		} while (seq != *t->seq);

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

	log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
	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;
	uint64_t seq = 0;
	int i;

	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++) {
		struct clock_entry *prev, *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
02bcaa8c	1	/*
f5cc024a	2	* Clock functions
02bcaa8c	3	*/
f5cc024a	4
02bcaa8c	5	#include <unistd.h>
c223da83	6	#include <math.h>
02bcaa8c	7	#include <sys/time.h>
03e20d68	8	#include <time.h>
02bcaa8c JA	9
02bcaa8c JA	10	#include "fio.h"
be4ecfdf	11	#include "smalloc.h"
02bcaa8c JA	12
02bcaa8c JA	13	#include "hash.h"
7d11f871	14	#include "os/os.h"
02bcaa8c	15
09a32402	16	#ifdef ARCH_HAVE_CPU_CLOCK
c223da83	17	static unsigned long cycles_per_usec;
71339117	18	static unsigned long inv_cycles_per_usec;
fa80feae	19	int tsc_reliable = 0;
09a32402	20	#endif
5d879392 JA	21
	22	struct tv_valid {
	23	struct timeval last_tv;
	24	int last_tv_valid;
93f0b09a	25	unsigned long last_cycles;
5d879392 JA	26	};
5d879392 JA	27	static pthread_key_t tv_tls_key;
02bcaa8c	28
16de1bf9	29	enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
fa80feae	30	int fio_clock_source_set = 0;
01423eae	31	enum fio_cs fio_clock_source_inited = CS_INVAL;
c223da83	32
02bcaa8c JA	33	#ifdef FIO_DEBUG_TIME
	34
	35	#define HASH_BITS 8
	36	#define HASH_SIZE (1 << HASH_BITS)
	37
01743ee1	38	static struct flist_head hash[HASH_SIZE];
02bcaa8c JA	39	static int gtod_inited;
	40
	41	struct gtod_log {
01743ee1	42	struct flist_head list;
02bcaa8c JA	43	void *caller;
	44	unsigned long calls;
	45	};
	46
	47	static struct gtod_log find_hash(void caller)
	48	{
	49	unsigned long h = hash_ptr(caller, HASH_BITS);
01743ee1	50	struct flist_head *entry;
02bcaa8c	51
01743ee1 JA	52	flist_for_each(entry, &hash[h]) {
	53	struct gtod_log *log = flist_entry(entry, struct gtod_log,
	54	list);
02bcaa8c JA	55
	56	if (log->caller == caller)
	57	return log;
	58	}
	59
	60	return NULL;
	61	}
	62
	63	static struct gtod_log find_log(void caller)
	64	{
	65	struct gtod_log *log = find_hash(caller);
	66
	67	if (!log) {
	68	unsigned long h;
	69
	70	log = malloc(sizeof(*log));
01743ee1	71	INIT_FLIST_HEAD(&log->list);
02bcaa8c JA	72	log->caller = caller;
	73	log->calls = 0;
	74
	75	h = hash_ptr(caller, HASH_BITS);
01743ee1	76	flist_add_tail(&log->list, &hash[h]);
02bcaa8c JA	77	}
	78
	79	return log;
	80	}
	81
	82	static void gtod_log_caller(void *caller)
	83	{
	84	if (gtod_inited) {
	85	struct gtod_log *log = find_log(caller);
	86
	87	log->calls++;
	88	}
	89	}
	90
	91	static void fio_exit fio_dump_gtod(void)
	92	{
	93	unsigned long total_calls = 0;
	94	int i;
	95
	96	for (i = 0; i < HASH_SIZE; i++) {
01743ee1	97	struct flist_head *entry;
02bcaa8c JA	98	struct gtod_log *log;
02bcaa8c JA	99
01743ee1 JA	100	flist_for_each(entry, &hash[i]) {
01743ee1 JA	101	log = flist_entry(entry, struct gtod_log, list);
02bcaa8c	102
5ec10eaa JA	103	printf("function %p, calls %lu\n", log->caller,
5ec10eaa JA	104	log->calls);
02bcaa8c JA	105	total_calls += log->calls;
	106	}
	107	}
	108
	109	printf("Total %lu gettimeofday\n", total_calls);
	110	}
	111
	112	static void fio_init gtod_init(void)
	113	{
	114	int i;
	115
	116	for (i = 0; i < HASH_SIZE; i++)
01743ee1	117	INIT_FLIST_HEAD(&hash[i]);
02bcaa8c JA	118
	119	gtod_inited = 1;
	120	}
	121
	122	#endif /* FIO_DEBUG_TIME */
	123
9ff1c070 JA	124	static int fill_clock_gettime(struct timespec *ts)
	125	{
	126	#ifdef FIO_HAVE_CLOCK_MONOTONIC
	127	return clock_gettime(CLOCK_MONOTONIC, ts);
	128	#else
	129	return clock_gettime(CLOCK_REALTIME, ts);
	130	#endif
	131	}
	132
1e97cce9	133	#ifdef FIO_DEBUG_TIME
02bcaa8c	134	void fio_gettime(struct timeval tp, void caller)
1e97cce9 JA	135	#else
	136	void fio_gettime(struct timeval tp, void fio_unused caller)
	137	#endif
02bcaa8c	138	{
93f0b09a	139	struct tv_valid *tv;
5d879392	140
02bcaa8c JA	141	#ifdef FIO_DEBUG_TIME
	142	if (!caller)
	143	caller = __builtin_return_address(0);
	144
	145	gtod_log_caller(caller);
02bcaa8c	146	#endif
be4ecfdf JA	147	if (fio_tv) {
	148	memcpy(tp, fio_tv, sizeof(*tp));
	149	return;
c223da83 JA	150	}
c223da83 JA	151
93f0b09a JA	152	tv = pthread_getspecific(tv_tls_key);
93f0b09a JA	153
c223da83 JA	154	switch (fio_clock_source) {
c223da83 JA	155	case CS_GTOD:
02bcaa8c	156	gettimeofday(tp, NULL);
c223da83 JA	157	break;
c223da83 JA	158	case CS_CGETTIME: {
02bcaa8c JA	159	struct timespec ts;
02bcaa8c JA	160
9ff1c070	161	if (fill_clock_gettime(&ts) < 0) {
c223da83 JA	162	log_err("fio: clock_gettime fails\n");
c223da83 JA	163	assert(0);
02bcaa8c JA	164	}
	165
	166	tp->tv_sec = ts.tv_sec;
	167	tp->tv_usec = ts.tv_nsec / 1000;
c223da83 JA	168	break;
	169	}
	170	#ifdef ARCH_HAVE_CPU_CLOCK
	171	case CS_CPUCLOCK: {
	172	unsigned long long usecs, t;
	173
	174	t = get_cpu_clock();
93f0b09a	175	if (tv && t < tv->last_cycles) {
c223da83	176	dprint(FD_TIME, "CPU clock going back in time\n");
93f0b09a JA	177	t = tv->last_cycles;
	178	} else if (tv)
	179	tv->last_cycles = t;
c223da83	180
71339117	181	usecs = (t * inv_cycles_per_usec) / 16777216UL;
c223da83 JA	182	tp->tv_sec = usecs / 1000000;
c223da83 JA	183	tp->tv_usec = usecs % 1000000;
c223da83 JA	184	break;
	185	}
	186	#endif
	187	default:
	188	log_err("fio: invalid clock source %d\n", fio_clock_source);
	189	break;
02bcaa8c	190	}
3e488920 JA	191
	192	/*
	193	* If Linux is using the tsc clock on non-synced processors,
	194	* sometimes time can appear to drift backwards. Fix that up.
	195	*/
93f0b09a JA	196	if (tv) {
	197	if (tv->last_tv_valid) {
	198	if (tp->tv_sec < tv->last_tv.tv_sec)
	199	tp->tv_sec = tv->last_tv.tv_sec;
	200	else if (tv->last_tv.tv_sec == tp->tv_sec &&
	201	tp->tv_usec < tv->last_tv.tv_usec)
	202	tp->tv_usec = tv->last_tv.tv_usec;
5d879392	203	}
93f0b09a JA	204	tv->last_tv_valid = 1;
93f0b09a JA	205	memcpy(&tv->last_tv, tp, sizeof(*tp));
3e488920	206	}
02bcaa8c	207	}
be4ecfdf	208
09a32402	209	#ifdef ARCH_HAVE_CPU_CLOCK
c223da83 JA	210	static unsigned long get_cycles_per_usec(void)
c223da83 JA	211	{
9ff1c070	212	struct timespec ts;
c223da83 JA	213	struct timeval s, e;
	214	unsigned long long c_s, c_e;
	215
9ff1c070 JA	216	fill_clock_gettime(&ts);
	217	s.tv_sec = ts.tv_sec;
	218	s.tv_usec = ts.tv_nsec / 1000;
	219
c223da83 JA	220	c_s = get_cpu_clock();
	221	do {
	222	unsigned long long elapsed;
	223
9ff1c070 JA	224	fill_clock_gettime(&ts);
	225	e.tv_sec = ts.tv_sec;
	226	e.tv_usec = ts.tv_nsec / 1000;
	227
c223da83	228	elapsed = utime_since(&s, &e);
486332e5	229	if (elapsed >= 1280) {
c223da83 JA	230	c_e = get_cpu_clock();
	231	break;
	232	}
	233	} while (1);
	234
c011000d	235	return (c_e - c_s + 127) >> 7;
c223da83 JA	236	}
c223da83 JA	237
fa80feae JA	238	#define NR_TIME_ITERS 50
fa80feae JA	239
09a32402	240	static void calibrate_cpu_clock(void)
c223da83 JA	241	{
c223da83 JA	242	double delta, mean, S;
fa80feae	243	unsigned long avg, cycles[NR_TIME_ITERS];
c223da83 JA	244	int i, samples;
c223da83 JA	245
c223da83 JA	246	cycles[0] = get_cycles_per_usec();
c223da83 JA	247	S = delta = mean = 0.0;
fa80feae	248	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	249	cycles[i] = get_cycles_per_usec();
	250	delta = cycles[i] - mean;
	251	if (delta) {
	252	mean += delta / (i + 1.0);
	253	S += delta * (cycles[i] - mean);
	254	}
	255	}
	256
fa80feae	257	S = sqrt(S / (NR_TIME_ITERS - 1.0));
c223da83 JA	258
c223da83 JA	259	samples = avg = 0;
fa80feae	260	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	261	double this = cycles[i];
c223da83 JA	262
03e20d68	263	if ((fmax(this, mean) - fmin(this, mean)) > S)
c223da83 JA	264	continue;
	265	samples++;
	266	avg += this;
	267	}
	268
fa80feae	269	S /= (double) NR_TIME_ITERS;
89db727d	270	mean /= 10.0;
c223da83	271
fa80feae	272	for (i = 0; i < NR_TIME_ITERS; i++)
c223da83 JA	273	dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
c223da83 JA	274
d7abad3d JA	275	avg /= samples;
d7abad3d JA	276	avg = (avg + 9) / 10;
c223da83 JA	277	dprint(FD_TIME, "avg: %lu\n", avg);
	278	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
	279
	280	cycles_per_usec = avg;
71339117 JA	281	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
71339117 JA	282	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
09a32402 JA	283	}
	284	#else
	285	static void calibrate_cpu_clock(void)
	286	{
	287	}
	288	#endif
	289
5d879392 JA	290	void fio_local_clock_init(int is_thread)
	291	{
	292	struct tv_valid *t;
	293
	294	t = calloc(sizeof(*t), 1);
	295	if (pthread_setspecific(tv_tls_key, t))
	296	log_err("fio: can't set TLS key\n");
	297	}
	298
	299	static void kill_tv_tls_key(void *data)
	300	{
	301	free(data);
	302	}
	303
09a32402 JA	304	void fio_clock_init(void)
09a32402 JA	305	{
01423eae JA	306	if (fio_clock_source == fio_clock_source_inited)
	307	return;
	308
5d879392 JA	309	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
	310	log_err("fio: can't create TLS key\n");
	311
01423eae	312	fio_clock_source_inited = fio_clock_source;
09a32402	313	calibrate_cpu_clock();
fa80feae JA	314
	315	/*
	316	* If the arch sets tsc_reliable != 0, then it must be good enough
	317	* to use as THE clock source. For x86 CPUs, this means the TSC
	318	* runs at a constant rate and is synced across CPU cores.
	319	*/
	320	if (tsc_reliable) {
	321	if (!fio_clock_source_set)
	322	fio_clock_source = CS_CPUCLOCK;
	323	} else if (fio_clock_source == CS_CPUCLOCK)
	324	log_info("fio: clocksource=cpu may not be reliable\n");
c223da83 JA	325	}
c223da83 JA	326
39ab7da2	327	unsigned long long utime_since(struct timeval s, struct timeval e)
be4ecfdf	328	{
39ab7da2 JA	329	long sec, usec;
	330	unsigned long long ret;
	331
	332	sec = e->tv_sec - s->tv_sec;
	333	usec = e->tv_usec - s->tv_usec;
	334	if (sec > 0 && usec < 0) {
	335	sec--;
	336	usec += 1000000;
	337	}
	338
	339	/*
	340	* time warp bug on some kernels?
	341	*/
	342	if (sec < 0 \|\| (sec == 0 && usec < 0))
	343	return 0;
	344
	345	ret = sec * 1000000ULL + usec;
	346
	347	return ret;
be4ecfdf JA	348	}
be4ecfdf JA	349
39ab7da2	350	unsigned long long utime_since_now(struct timeval *s)
be4ecfdf	351	{
39ab7da2 JA	352	struct timeval t;
	353
	354	fio_gettime(&t, NULL);
	355	return utime_since(s, &t);
be4ecfdf	356	}
783a3eb1	357
39ab7da2	358	unsigned long mtime_since(struct timeval s, struct timeval e)
783a3eb1	359	{
39ab7da2	360	long sec, usec, ret;
783a3eb1	361
39ab7da2 JA	362	sec = e->tv_sec - s->tv_sec;
	363	usec = e->tv_usec - s->tv_usec;
	364	if (sec > 0 && usec < 0) {
	365	sec--;
	366	usec += 1000000;
783a3eb1 JA	367	}
783a3eb1 JA	368
39ab7da2 JA	369	if (sec < 0 \|\| (sec == 0 && usec < 0))
	370	return 0;
	371
	372	sec *= 1000UL;
	373	usec /= 1000UL;
	374	ret = sec + usec;
	375
	376	return ret;
783a3eb1 JA	377	}
783a3eb1 JA	378
39ab7da2	379	unsigned long mtime_since_now(struct timeval *s)
783a3eb1	380	{
39ab7da2 JA	381	struct timeval t;
39ab7da2 JA	382	void *p = __builtin_return_address(0);
783a3eb1	383
39ab7da2 JA	384	fio_gettime(&t, p);
	385	return mtime_since(s, &t);
	386	}
783a3eb1	387
39ab7da2 JA	388	unsigned long time_since_now(struct timeval *s)
	389	{
	390	return mtime_since_now(s) / 1000;
783a3eb1	391	}
7d11f871 JA	392
	393	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)
	394
	395	#define CLOCK_ENTRIES 100000
	396
	397	struct clock_entry {
	398	unsigned long seq;
	399	unsigned long tsc;
	400	unsigned long cpu;
	401	};
	402
	403	struct clock_thread {
	404	pthread_t thread;
	405	int cpu;
	406	pthread_mutex_t lock;
	407	pthread_mutex_t started;
	408	uint64_t *seq;
	409	struct clock_entry *entries;
	410	};
	411
	412	static inline uint64_t atomic64_inc_return(uint64_t *seq)
	413	{
	414	return 1 + __sync_fetch_and_add(seq, 1);
	415	}
	416
	417	static void clock_thread_fn(void data)
	418	{
	419	struct clock_thread *t = data;
	420	struct clock_entry *c;
	421	os_cpu_mask_t cpu_mask;
	422	int i;
	423
	424	memset(&cpu_mask, 0, sizeof(cpu_mask));
	425	fio_cpu_set(&cpu_mask, t->cpu);
	426
	427	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
	428	log_err("clock setaffinity failed\n");
	429	return (void *) 1;
	430	}
	431
7d11f871	432	pthread_mutex_lock(&t->lock);
b9b3498e	433	pthread_mutex_unlock(&t->started);
7d11f871 JA	434
	435	c = &t->entries[0];
	436	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
	437	uint64_t seq, tsc;
	438
	439	c->cpu = t->cpu;
	440	do {
	441	seq = atomic64_inc_return(t->seq);
	442	tsc = get_cpu_clock();
	443	} while (seq != *t->seq);
	444
	445	c->seq = seq;
	446	c->tsc = tsc;
	447	}
	448
	449	log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
	450	return NULL;
	451	}
	452
	453	static int clock_cmp(const void p1, const void p2)
	454	{
	455	const struct clock_entry *c1 = p1;
	456	const struct clock_entry *c2 = p2;
	457
b9b3498e JA	458	if (c1->seq == c2->seq)
	459	log_err("cs: bug in atomic sequence!\n");
	460
7d11f871 JA	461	return c1->seq - c2->seq;
	462	}
	463
	464	int fio_monotonic_clocktest(void)
	465	{
	466	struct clock_thread *threads;
	467	unsigned int nr_cpus = cpus_online();
	468	struct clock_entry *entries;
	469	unsigned long tentries, failed;
	470	uint64_t seq = 0;
	471	int i;
	472
4f1d43c2 JA	473	fio_debug \|= 1U << FD_TIME;
	474	calibrate_cpu_clock();
	475	fio_debug &= ~(1U << FD_TIME);
	476
7d11f871 JA	477	threads = malloc(nr_cpus * sizeof(struct clock_thread));
	478	tentries = CLOCK_ENTRIES * nr_cpus;
	479	entries = malloc(tentries * sizeof(struct clock_entry));
	480
	481	log_info("cs: Testing %u CPUs\n", nr_cpus);
	482
	483	for (i = 0; i < nr_cpus; i++) {
	484	struct clock_thread *t = &threads[i];
	485
	486	t->cpu = i;
	487	t->seq = &seq;
	488	t->entries = &entries[i * CLOCK_ENTRIES];
	489	pthread_mutex_init(&t->lock, NULL);
	490	pthread_mutex_init(&t->started, NULL);
	491	pthread_mutex_lock(&t->lock);
	492	pthread_create(&t->thread, NULL, clock_thread_fn, t);
	493	}
	494
	495	for (i = 0; i < nr_cpus; i++) {
	496	struct clock_thread *t = &threads[i];
	497
	498	pthread_mutex_lock(&t->started);
	499	}
	500
	501	for (i = 0; i < nr_cpus; i++) {
	502	struct clock_thread *t = &threads[i];
	503
	504	pthread_mutex_unlock(&t->lock);
	505	}
	506
	507	for (failed = i = 0; i < nr_cpus; i++) {
	508	struct clock_thread *t = &threads[i];
	509	void *ret;
	510
	511	pthread_join(t->thread, &ret);
	512	if (ret)
	513	failed++;
	514	}
	515	free(threads);
	516
	517	if (failed) {
	518	log_err("Clocksource test: %u threads failed\n", failed);
	519	goto err;
	520	}
	521
	522	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	523
	524	for (failed = i = 0; i < tentries; i++) {
	525	struct clock_entry prev, this = &entries[i];
	526
	527	if (!i) {
	528	prev = this;
	529	continue;
	530	}
	531
	532	if (prev->tsc > this->tsc) {
	533	uint64_t diff = prev->tsc - this->tsc;
	534
	535	log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
	536	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
	537	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
	538	failed++;
	539	}
	540
541	prev = this;
542	}
543
544	if (failed)
545	log_info("cs: Failed: %lu\n", failed);
546	else
547	log_info("cs: Pass!\n");
548
549	err:
550	free(entries);
551	return !!failed;
552	}
553
554	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
555
556	int fio_monotonic_clocktest(void)
557	{
558	log_info("cs: current platform does not support CPU clocks\n");
559	return 0;
560	}
561
562	#endif