[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"

#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
static unsigned long cycles_per_usec;
static unsigned long inv_cycles_per_usec;
#endif
int tsc_reliable = 0;

struct tv_valid {
	uint64_t last_cycles;
	int last_tv_valid;
	int warned;
};
#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;
static 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 (t < tv->last_cycles && tv->last_tv_valid &&
		    !tv->warned) {
			log_err("fio: CPU clock going back in time\n");
			tv->warned = 1;
		}

		tv->last_cycles = t;
		tv->last_tv_valid = 1;
#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
		usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
#else
		usecs = (t * inv_cycles_per_usec) / 16777216UL;
#endif
		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;
	}
}

#ifdef FIO_DEBUG_TIME
void fio_gettime(struct timeval *tp, void *caller)
#else
void fio_gettime(struct timeval *tp, void fio_unused *caller)
#endif
{
#ifdef FIO_DEBUG_TIME
	if (!caller)
		caller = __builtin_return_address(0);

	gtod_log_caller(caller);
#endif
	if (fio_unlikely(fio_gettime_offload(tp)))
		return;

	__fio_gettime(tp);
}

#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
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)
{
#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
	return 0;
#else
	return 1;
#endif
}
#endif // ARCH_HAVE_CPU_CLOCK

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

	t = calloc(1, sizeof(*t));
	if (pthread_setspecific(tv_tls_key, t)) {
		log_err("fio: can't set TLS key\n");
		assert(0);
	}
}

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(const struct timeval *s, const 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(const struct timeval *s)
{
	struct timeval t;

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

uint64_t mtime_since(const struct timeval *s, const 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(const 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(const 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 *cthreads;
	unsigned int nr_cpus = cpus_online();
	struct clock_entry *entries;
	unsigned long tentries, failed = 0;
	struct clock_entry *prev, *this;
	uint32_t seq = 0;
	unsigned int i;

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

#ifdef FIO_INC_DEBUG
	fio_debug |= 1U << FD_TIME;
#endif
	calibrate_cpu_clock();
#ifdef FIO_INC_DEBUG
	fio_debug &= ~(1U << FD_TIME);
#endif

	cthreads = 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 = &cthreads[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);
		if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
			failed++;
			nr_cpus = i;
			break;
		}
	}

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

		pthread_mutex_lock(&t->started);
	}

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

		pthread_mutex_unlock(&t->lock);
	}

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

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

	if (failed) {
		log_err("Clocksource test: %lu 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=%llu):\n",
						(unsigned long long) diff);
			log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
			log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) 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	#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
	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	uint64_t last_cycles;
	24	int last_tv_valid;
	25	int warned;
	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	static 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 (t < tv->last_cycles && tv->last_tv_valid &&
	175	!tv->warned) {
	176	log_err("fio: CPU clock going back in time\n");
	177	tv->warned = 1;
	178	}
	179
	180	tv->last_cycles = t;
	181	tv->last_tv_valid = 1;
	182	#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
	183	usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
	184	#else
	185	usecs = (t * inv_cycles_per_usec) / 16777216UL;
	186	#endif
	187	tp->tv_sec = usecs / 1000000;
	188	tp->tv_usec = usecs % 1000000;
	189	break;
	190	}
	191	#endif
	192	default:
	193	log_err("fio: invalid clock source %d\n", fio_clock_source);
	194	break;
	195	}
	196	}
	197
	198	#ifdef FIO_DEBUG_TIME
	199	void fio_gettime(struct timeval tp, void caller)
	200	#else
	201	void fio_gettime(struct timeval tp, void fio_unused caller)
	202	#endif
	203	{
	204	#ifdef FIO_DEBUG_TIME
	205	if (!caller)
	206	caller = __builtin_return_address(0);
	207
	208	gtod_log_caller(caller);
	209	#endif
	210	if (fio_unlikely(fio_gettime_offload(tp)))
	211	return;
	212
	213	__fio_gettime(tp);
	214	}
	215
	216	#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
	217	static unsigned long get_cycles_per_usec(void)
	218	{
	219	struct timeval s, e;
	220	uint64_t c_s, c_e;
	221	enum fio_cs old_cs = fio_clock_source;
	222
	223	#ifdef CONFIG_CLOCK_GETTIME
	224	fio_clock_source = CS_CGETTIME;
	225	#else
	226	fio_clock_source = CS_GTOD;
	227	#endif
	228	__fio_gettime(&s);
	229
	230	c_s = get_cpu_clock();
	231	do {
	232	uint64_t elapsed;
	233
	234	__fio_gettime(&e);
	235
	236	elapsed = utime_since(&s, &e);
	237	if (elapsed >= 1280) {
	238	c_e = get_cpu_clock();
	239	break;
	240	}
	241	} while (1);
	242
	243	fio_clock_source = old_cs;
	244	return (c_e - c_s + 127) >> 7;
	245	}
	246
	247	#define NR_TIME_ITERS 50
	248
	249	static int calibrate_cpu_clock(void)
	250	{
	251	double delta, mean, S;
	252	uint64_t avg, cycles[NR_TIME_ITERS];
	253	int i, samples;
	254
	255	cycles[0] = get_cycles_per_usec();
	256	S = delta = mean = 0.0;
	257	for (i = 0; i < NR_TIME_ITERS; i++) {
	258	cycles[i] = get_cycles_per_usec();
	259	delta = cycles[i] - mean;
	260	if (delta) {
	261	mean += delta / (i + 1.0);
	262	S += delta * (cycles[i] - mean);
	263	}
	264	}
	265
	266	/*
	267	* The most common platform clock breakage is returning zero
	268	* indefinitely. Check for that and return failure.
	269	*/
	270	if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
	271	return 1;
	272
	273	S = sqrt(S / (NR_TIME_ITERS - 1.0));
	274
	275	samples = avg = 0;
	276	for (i = 0; i < NR_TIME_ITERS; i++) {
	277	double this = cycles[i];
	278
	279	if ((fmax(this, mean) - fmin(this, mean)) > S)
	280	continue;
	281	samples++;
	282	avg += this;
	283	}
	284
	285	S /= (double) NR_TIME_ITERS;
	286	mean /= 10.0;
	287
	288	for (i = 0; i < NR_TIME_ITERS; i++)
	289	dprint(FD_TIME, "cycles[%d]=%llu\n", i,
	290	(unsigned long long) cycles[i] / 10);
	291
	292	avg /= samples;
	293	avg = (avg + 5) / 10;
	294	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
	295	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
	296
	297	cycles_per_usec = avg;
	298	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
	299	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
	300	return 0;
	301	}
	302	#else
	303	static int calibrate_cpu_clock(void)
	304	{
	305	#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
	306	return 0;
	307	#else
	308	return 1;
	309	#endif
	310	}
	311	#endif // ARCH_HAVE_CPU_CLOCK
	312
	313	#ifndef CONFIG_TLS_THREAD
	314	void fio_local_clock_init(int is_thread)
	315	{
	316	struct tv_valid *t;
	317
	318	t = calloc(1, sizeof(*t));
	319	if (pthread_setspecific(tv_tls_key, t)) {
	320	log_err("fio: can't set TLS key\n");
	321	assert(0);
	322	}
	323	}
	324
	325	static void kill_tv_tls_key(void *data)
	326	{
	327	free(data);
	328	}
	329	#else
	330	void fio_local_clock_init(int is_thread)
	331	{
	332	}
	333	#endif
	334
	335	void fio_clock_init(void)
	336	{
	337	if (fio_clock_source == fio_clock_source_inited)
	338	return;
	339
	340	#ifndef CONFIG_TLS_THREAD
	341	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
	342	log_err("fio: can't create TLS key\n");
	343	#endif
	344
	345	fio_clock_source_inited = fio_clock_source;
	346
	347	if (calibrate_cpu_clock())
	348	tsc_reliable = 0;
	349
	350	/*
	351	* If the arch sets tsc_reliable != 0, then it must be good enough
	352	* to use as THE clock source. For x86 CPUs, this means the TSC
	353	* runs at a constant rate and is synced across CPU cores.
	354	*/
	355	if (tsc_reliable) {
	356	if (!fio_clock_source_set)
	357	fio_clock_source = CS_CPUCLOCK;
	358	} else if (fio_clock_source == CS_CPUCLOCK)
	359	log_info("fio: clocksource=cpu may not be reliable\n");
	360	}
	361
	362	uint64_t utime_since(const struct timeval s, const struct timeval e)
	363	{
	364	long sec, usec;
	365	uint64_t ret;
	366
	367	sec = e->tv_sec - s->tv_sec;
	368	usec = e->tv_usec - s->tv_usec;
	369	if (sec > 0 && usec < 0) {
	370	sec--;
	371	usec += 1000000;
	372	}
	373
	374	/*
	375	* time warp bug on some kernels?
	376	*/
	377	if (sec < 0 \|\| (sec == 0 && usec < 0))
	378	return 0;
	379
	380	ret = sec * 1000000ULL + usec;
	381
	382	return ret;
	383	}
	384
	385	uint64_t utime_since_now(const struct timeval *s)
	386	{
	387	struct timeval t;
	388
	389	fio_gettime(&t, NULL);
	390	return utime_since(s, &t);
	391	}
	392
	393	uint64_t mtime_since(const struct timeval s, const struct timeval e)
	394	{
	395	long sec, usec, ret;
	396
	397	sec = e->tv_sec - s->tv_sec;
	398	usec = e->tv_usec - s->tv_usec;
	399	if (sec > 0 && usec < 0) {
	400	sec--;
	401	usec += 1000000;
	402	}
	403
	404	if (sec < 0 \|\| (sec == 0 && usec < 0))
	405	return 0;
	406
	407	sec *= 1000UL;
	408	usec /= 1000UL;
	409	ret = sec + usec;
	410
	411	return ret;
	412	}
	413
	414	uint64_t mtime_since_now(const struct timeval *s)
	415	{
	416	struct timeval t;
	417	void *p = __builtin_return_address(0);
	418
	419	fio_gettime(&t, p);
	420	return mtime_since(s, &t);
	421	}
	422
	423	uint64_t time_since_now(const struct timeval *s)
	424	{
	425	return mtime_since_now(s) / 1000;
	426	}
	427
	428	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
	429	defined(CONFIG_SFAA)
	430
	431	#define CLOCK_ENTRIES 100000
	432
	433	struct clock_entry {
	434	uint32_t seq;
	435	uint32_t cpu;
	436	uint64_t tsc;
	437	};
	438
	439	struct clock_thread {
	440	pthread_t thread;
	441	int cpu;
	442	pthread_mutex_t lock;
	443	pthread_mutex_t started;
	444	uint32_t *seq;
	445	struct clock_entry *entries;
	446	};
	447
	448	static inline uint32_t atomic32_inc_return(uint32_t *seq)
	449	{
	450	return 1 + __sync_fetch_and_add(seq, 1);
	451	}
	452
	453	static void clock_thread_fn(void data)
	454	{
	455	struct clock_thread *t = data;
	456	struct clock_entry *c;
	457	os_cpu_mask_t cpu_mask;
	458	uint32_t last_seq;
	459	int i;
	460
	461	memset(&cpu_mask, 0, sizeof(cpu_mask));
	462	fio_cpu_set(&cpu_mask, t->cpu);
	463
	464	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
	465	log_err("clock setaffinity failed\n");
	466	return (void *) 1;
	467	}
	468
	469	pthread_mutex_lock(&t->lock);
	470	pthread_mutex_unlock(&t->started);
	471
	472	last_seq = 0;
	473	c = &t->entries[0];
	474	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
	475	uint32_t seq;
	476	uint64_t tsc;
	477
	478	c->cpu = t->cpu;
	479	do {
	480	seq = atomic32_inc_return(t->seq);
	481	if (seq < last_seq)
	482	break;
	483	tsc = get_cpu_clock();
	484	} while (seq != *t->seq);
	485
	486	c->seq = seq;
	487	c->tsc = tsc;
	488	}
	489
	490	log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
	491	(unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
	492
	493	/*
	494	* The most common platform clock breakage is returning zero
	495	* indefinitely. Check for that and return failure.
	496	*/
	497	if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
	498	return (void *) 1;
	499
	500	return NULL;
	501	}
	502
	503	static int clock_cmp(const void p1, const void p2)
	504	{
	505	const struct clock_entry *c1 = p1;
	506	const struct clock_entry *c2 = p2;
	507
	508	if (c1->seq == c2->seq)
	509	log_err("cs: bug in atomic sequence!\n");
	510
	511	return c1->seq - c2->seq;
	512	}
	513
	514	int fio_monotonic_clocktest(void)
	515	{
	516	struct clock_thread *cthreads;
	517	unsigned int nr_cpus = cpus_online();
	518	struct clock_entry *entries;
	519	unsigned long tentries, failed = 0;
	520	struct clock_entry prev, this;
	521	uint32_t seq = 0;
	522	unsigned int i;
	523
	524	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
	525
	526	#ifdef FIO_INC_DEBUG
	527	fio_debug \|= 1U << FD_TIME;
	528	#endif
	529	calibrate_cpu_clock();
	530	#ifdef FIO_INC_DEBUG
	531	fio_debug &= ~(1U << FD_TIME);
	532	#endif
	533
	534	cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
	535	tentries = CLOCK_ENTRIES * nr_cpus;
	536	entries = malloc(tentries * sizeof(struct clock_entry));
	537
	538	log_info("cs: Testing %u CPUs\n", nr_cpus);
	539
	540	for (i = 0; i < nr_cpus; i++) {
	541	struct clock_thread *t = &cthreads[i];
	542
	543	t->cpu = i;
	544	t->seq = &seq;
	545	t->entries = &entries[i * CLOCK_ENTRIES];
	546	pthread_mutex_init(&t->lock, NULL);
	547	pthread_mutex_init(&t->started, NULL);
	548	pthread_mutex_lock(&t->lock);
	549	if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
	550	failed++;
	551	nr_cpus = i;
	552	break;
	553	}
	554	}
	555
	556	for (i = 0; i < nr_cpus; i++) {
	557	struct clock_thread *t = &cthreads[i];
	558
	559	pthread_mutex_lock(&t->started);
	560	}
	561
	562	for (i = 0; i < nr_cpus; i++) {
	563	struct clock_thread *t = &cthreads[i];
	564
	565	pthread_mutex_unlock(&t->lock);
	566	}
	567
	568	for (i = 0; i < nr_cpus; i++) {
	569	struct clock_thread *t = &cthreads[i];
	570	void *ret;
	571
	572	pthread_join(t->thread, &ret);
	573	if (ret)
	574	failed++;
	575	}
	576	free(cthreads);
	577
	578	if (failed) {
	579	log_err("Clocksource test: %lu threads failed\n", failed);
	580	goto err;
	581	}
	582
	583	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	584
	585	for (failed = i = 0; i < tentries; i++) {
	586	this = &entries[i];
	587
	588	if (!i) {
	589	prev = this;
	590	continue;
	591	}
	592
	593	if (prev->tsc > this->tsc) {
	594	uint64_t diff = prev->tsc - this->tsc;
	595
	596	log_info("cs: CPU clock mismatch (diff=%llu):\n",
	597	(unsigned long long) diff);
	598	log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
	599	log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
	600	failed++;
	601	}
	602
	603	prev = this;
	604	}
	605
	606	if (failed)
	607	log_info("cs: Failed: %lu\n", failed);
	608	else
	609	log_info("cs: Pass!\n");
	610
	611	err:
	612	free(entries);
	613	return !!failed;
	614	}
	615
	616	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
	617
	618	int fio_monotonic_clocktest(void)
	619	{
	620	log_info("cs: current platform does not support CPU clocks\n");
	621	return 0;
	622	}
	623
	624	#endif