[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)
#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
static unsigned long cycles_per_usec;
static unsigned long inv_cycles_per_usec;
static uint64_t max_cycles_for_mult;
#endif
#ifdef ARCH_CPU_CLOCK_WRAPS
static unsigned long long cycles_start, cycles_wrap;
#endif
#endif
int tsc_reliable = 0;

struct tv_valid {
	uint64_t last_cycles;
	int last_tv_valid;
	int warned;
};
#ifdef ARCH_HAVE_CPU_CLOCK
#ifdef CONFIG_TLS_THREAD
static __thread struct tv_valid static_tv_valid;
#else
static pthread_key_t tv_tls_key;
#endif
#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 void inc_caller(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]);
	}

	log->calls++;
}

static void gtod_log_caller(void *caller)
{
	if (gtod_inited)
		inc_caller(caller);
}

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)
{
#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
	return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
#elif defined(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)
{
	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;
		struct tv_valid *tv;

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

		t = get_cpu_clock();
#ifdef ARCH_CPU_CLOCK_WRAPS
		if (t < cycles_start && !cycles_wrap)
			cycles_wrap = 1;
		else if (cycles_wrap && t >= cycles_start && !tv->warned) {
			log_err("fio: double CPU clock wrap\n");
			tv->warned = 1;
		}

		t -= cycles_start;
#endif
		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
		if (t < max_cycles_for_mult)
			usecs = (t * inv_cycles_per_usec) / 16777216UL;
		else
			usecs = t / cycles_per_usec;
#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 minc, maxc, 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));

	minc = -1ULL;
	maxc = samples = avg = 0;
	for (i = 0; i < NR_TIME_ITERS; i++) {
		double this = cycles[i];

		minc = min(cycles[i], minc);
		maxc = max(cycles[i], maxc);

		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;
	minc /= 10;
	maxc /= 10;
	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
	dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
			(unsigned long long) minc,
			(unsigned long long) maxc, mean, S);

	cycles_per_usec = avg;
	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
	max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
#ifdef ARCH_CPU_CLOCK_WRAPS
	cycles_start = get_cpu_clock();
	dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
#endif
	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_monotonic_clocktest(0))
			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_DEBUG	100000
#define CLOCK_ENTRIES_TEST	10000

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

struct clock_thread {
	pthread_t thread;
	int cpu;
	int debug;
	pthread_mutex_t lock;
	pthread_mutex_t started;
	unsigned long nr_entries;
	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;

	if (fio_cpuset_init(&cpu_mask)) {
		int __err = errno;

		log_err("clock cpuset init failed: %s\n", strerror(__err));
		goto err_out;
	}

	fio_cpu_set(&cpu_mask, t->cpu);

	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
		int __err = errno;

		log_err("clock setaffinity failed: %s\n", strerror(__err));
		goto err;
	}

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

	last_seq = 0;
	c = &t->entries[0];
	for (i = 0; i < t->nr_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;
	}

	if (t->debug) {
		unsigned long long clocks;

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

	/*
	 * 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)
		goto err;

	fio_cpuset_exit(&cpu_mask);
	return NULL;
err:
	fio_cpuset_exit(&cpu_mask);
err_out:
	return (void *) 1;
}

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

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

#ifdef FIO_INC_DEBUG
		fio_debug |= 1U << FD_TIME;
#endif
		nr_entries = CLOCK_ENTRIES_DEBUG;
	} else
		nr_entries = CLOCK_ENTRIES_TEST;

	calibrate_cpu_clock();

	if (debug) {
#ifdef FIO_INC_DEBUG
		fio_debug &= ~(1U << FD_TIME);
#endif
	}

	cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
	tentries = nr_entries * nr_cpus;
	entries = malloc(tentries * sizeof(struct clock_entry));

	if (debug)
		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->debug = debug;
		t->seq = &seq;
		t->nr_entries = nr_entries;
		t->entries = &entries[i * nr_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) {
		if (debug)
			log_err("Clocksource test: %lu threads failed\n", failed);
		goto err;
	}

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

	/* silence silly gcc */
	prev = NULL;
	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;

			if (!debug) {
				failed++;
				break;
			}

			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 (debug) {
		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(int debug)
{
	if (debug)
		log_info("cs: current platform does not support CPU clocks\n");
	return 1;
}

#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
dac499a0 AV	16	#if defined(ARCH_HAVE_CPU_CLOCK)
dac499a0 AV	17	#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
c223da83	18	static unsigned long cycles_per_usec;
71339117	19	static unsigned long inv_cycles_per_usec;
b3fa625b	20	static uint64_t max_cycles_for_mult;
96170421 CE	21	#endif
96170421 CE	22	#ifdef ARCH_CPU_CLOCK_WRAPS
73df3e07	23	static unsigned long long cycles_start, cycles_wrap;
09a32402	24	#endif
dac499a0	25	#endif
4de98eb0	26	int tsc_reliable = 0;
5d879392 JA	27
5d879392 JA	28	struct tv_valid {
ba458c2f	29	uint64_t last_cycles;
02dcf81c JA	30	int last_tv_valid;
02dcf81c JA	31	int warned;
5d879392	32	};
03be65c9	33	#ifdef ARCH_HAVE_CPU_CLOCK
67bf9823	34	#ifdef CONFIG_TLS_THREAD
b4ea84da	35	static __thread struct tv_valid static_tv_valid;
67bf9823	36	#else
5d879392	37	static pthread_key_t tv_tls_key;
67bf9823	38	#endif
03be65c9	39	#endif
02bcaa8c	40
16de1bf9	41	enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
fa80feae	42	int fio_clock_source_set = 0;
10aa136b	43	static enum fio_cs fio_clock_source_inited = CS_INVAL;
c223da83	44
02bcaa8c JA	45	#ifdef FIO_DEBUG_TIME
	46
	47	#define HASH_BITS 8
	48	#define HASH_SIZE (1 << HASH_BITS)
	49
01743ee1	50	static struct flist_head hash[HASH_SIZE];
02bcaa8c JA	51	static int gtod_inited;
	52
	53	struct gtod_log {
01743ee1	54	struct flist_head list;
02bcaa8c JA	55	void *caller;
	56	unsigned long calls;
	57	};
	58
	59	static struct gtod_log find_hash(void caller)
	60	{
	61	unsigned long h = hash_ptr(caller, HASH_BITS);
01743ee1	62	struct flist_head *entry;
02bcaa8c	63
01743ee1 JA	64	flist_for_each(entry, &hash[h]) {
	65	struct gtod_log *log = flist_entry(entry, struct gtod_log,
	66	list);
02bcaa8c JA	67
	68	if (log->caller == caller)
	69	return log;
	70	}
	71
	72	return NULL;
	73	}
	74
d5e16441	75	static void inc_caller(void *caller)
02bcaa8c JA	76	{
	77	struct gtod_log *log = find_hash(caller);
	78
	79	if (!log) {
	80	unsigned long h;
	81
	82	log = malloc(sizeof(*log));
01743ee1	83	INIT_FLIST_HEAD(&log->list);
02bcaa8c JA	84	log->caller = caller;
	85	log->calls = 0;
	86
	87	h = hash_ptr(caller, HASH_BITS);
01743ee1	88	flist_add_tail(&log->list, &hash[h]);
02bcaa8c JA	89	}
02bcaa8c JA	90
d5e16441	91	log->calls++;
02bcaa8c JA	92	}
	93
	94	static void gtod_log_caller(void *caller)
	95	{
d5e16441 JA	96	if (gtod_inited)
d5e16441 JA	97	inc_caller(caller);
02bcaa8c JA	98	}
	99
	100	static void fio_exit fio_dump_gtod(void)
	101	{
	102	unsigned long total_calls = 0;
	103	int i;
	104
	105	for (i = 0; i < HASH_SIZE; i++) {
01743ee1	106	struct flist_head *entry;
02bcaa8c JA	107	struct gtod_log *log;
02bcaa8c JA	108
01743ee1 JA	109	flist_for_each(entry, &hash[i]) {
01743ee1 JA	110	log = flist_entry(entry, struct gtod_log, list);
02bcaa8c	111
5ec10eaa JA	112	printf("function %p, calls %lu\n", log->caller,
5ec10eaa JA	113	log->calls);
02bcaa8c JA	114	total_calls += log->calls;
	115	}
	116	}
	117
	118	printf("Total %lu gettimeofday\n", total_calls);
	119	}
	120
	121	static void fio_init gtod_init(void)
	122	{
	123	int i;
	124
	125	for (i = 0; i < HASH_SIZE; i++)
01743ee1	126	INIT_FLIST_HEAD(&hash[i]);
02bcaa8c JA	127
	128	gtod_inited = 1;
	129	}
	130
	131	#endif /* FIO_DEBUG_TIME */
	132
67bf9823	133	#ifdef CONFIG_CLOCK_GETTIME
9ff1c070 JA	134	static int fill_clock_gettime(struct timespec *ts)
9ff1c070 JA	135	{
c544f604 SN	136	#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
	137	return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
	138	#elif defined(CONFIG_CLOCK_MONOTONIC)
9ff1c070 JA	139	return clock_gettime(CLOCK_MONOTONIC, ts);
	140	#else
	141	return clock_gettime(CLOCK_REALTIME, ts);
	142	#endif
	143	}
1e97cce9	144	#endif
67bf9823	145
9eb271b9	146	static void __fio_gettime(struct timeval *tp)
02bcaa8c	147	{
c223da83	148	switch (fio_clock_source) {
67bf9823	149	#ifdef CONFIG_GETTIMEOFDAY
c223da83	150	case CS_GTOD:
02bcaa8c	151	gettimeofday(tp, NULL);
c223da83	152	break;
67bf9823 JA	153	#endif
67bf9823 JA	154	#ifdef CONFIG_CLOCK_GETTIME
c223da83	155	case CS_CGETTIME: {
02bcaa8c JA	156	struct timespec ts;
02bcaa8c JA	157
9ff1c070	158	if (fill_clock_gettime(&ts) < 0) {
c223da83 JA	159	log_err("fio: clock_gettime fails\n");
c223da83 JA	160	assert(0);
02bcaa8c JA	161	}
	162
	163	tp->tv_sec = ts.tv_sec;
	164	tp->tv_usec = ts.tv_nsec / 1000;
c223da83 JA	165	break;
c223da83 JA	166	}
67bf9823	167	#endif
c223da83 JA	168	#ifdef ARCH_HAVE_CPU_CLOCK
c223da83 JA	169	case CS_CPUCLOCK: {
ba458c2f	170	uint64_t usecs, t;
03be65c9 JA	171	struct tv_valid *tv;
	172
	173	#ifdef CONFIG_TLS_THREAD
	174	tv = &static_tv_valid;
	175	#else
	176	tv = pthread_getspecific(tv_tls_key);
	177	#endif
c223da83 JA	178
c223da83 JA	179	t = get_cpu_clock();
96170421	180	#ifdef ARCH_CPU_CLOCK_WRAPS
73df3e07 JA	181	if (t < cycles_start && !cycles_wrap)
73df3e07 JA	182	cycles_wrap = 1;
03be65c9 JA	183	else if (cycles_wrap && t >= cycles_start && !tv->warned) {
	184	log_err("fio: double CPU clock wrap\n");
	185	tv->warned = 1;
02dcf81c	186	}
c223da83	187
73df3e07	188	t -= cycles_start;
96170421	189	#endif
9eb271b9 JA	190	tv->last_cycles = t;
9eb271b9 JA	191	tv->last_tv_valid = 1;
919e789d CE	192	#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
	193	usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
	194	#else
b3fa625b JA	195	if (t < max_cycles_for_mult)
	196	usecs = (t * inv_cycles_per_usec) / 16777216UL;
	197	else
	198	usecs = t / cycles_per_usec;
919e789d	199	#endif
c223da83 JA	200	tp->tv_sec = usecs / 1000000;
c223da83 JA	201	tp->tv_usec = usecs % 1000000;
c223da83 JA	202	break;
	203	}
	204	#endif
	205	default:
	206	log_err("fio: invalid clock source %d\n", fio_clock_source);
	207	break;
02bcaa8c	208	}
67bf9823 JA	209	}
	210
	211	#ifdef FIO_DEBUG_TIME
	212	void fio_gettime(struct timeval tp, void caller)
	213	#else
	214	void fio_gettime(struct timeval tp, void fio_unused caller)
	215	#endif
	216	{
67bf9823 JA	217	#ifdef FIO_DEBUG_TIME
	218	if (!caller)
	219	caller = __builtin_return_address(0);
	220
	221	gtod_log_caller(caller);
	222	#endif
27325ed5	223	if (fio_unlikely(fio_gettime_offload(tp)))
67bf9823	224	return;
67bf9823	225
9eb271b9	226	__fio_gettime(tp);
02bcaa8c	227	}
be4ecfdf	228
919e789d	229	#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
c223da83 JA	230	static unsigned long get_cycles_per_usec(void)
	231	{
	232	struct timeval s, e;
ba458c2f	233	uint64_t c_s, c_e;
67bf9823	234	enum fio_cs old_cs = fio_clock_source;
c223da83	235
67bf9823 JA	236	#ifdef CONFIG_CLOCK_GETTIME
	237	fio_clock_source = CS_CGETTIME;
	238	#else
	239	fio_clock_source = CS_GTOD;
	240	#endif
	241	__fio_gettime(&s);
9ff1c070	242
c223da83 JA	243	c_s = get_cpu_clock();
c223da83 JA	244	do {
ba458c2f	245	uint64_t elapsed;
c223da83	246
67bf9823	247	__fio_gettime(&e);
9ff1c070	248
c223da83	249	elapsed = utime_since(&s, &e);
486332e5	250	if (elapsed >= 1280) {
c223da83 JA	251	c_e = get_cpu_clock();
	252	break;
	253	}
	254	} while (1);
	255
67bf9823	256	fio_clock_source = old_cs;
c011000d	257	return (c_e - c_s + 127) >> 7;
c223da83 JA	258	}
c223da83 JA	259
fa80feae JA	260	#define NR_TIME_ITERS 50
fa80feae JA	261
e259879e	262	static int calibrate_cpu_clock(void)
c223da83 JA	263	{
c223da83 JA	264	double delta, mean, S;
e51a6629	265	uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
c223da83 JA	266	int i, samples;
c223da83 JA	267
c223da83 JA	268	cycles[0] = get_cycles_per_usec();
c223da83 JA	269	S = delta = mean = 0.0;
fa80feae	270	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	271	cycles[i] = get_cycles_per_usec();
	272	delta = cycles[i] - mean;
	273	if (delta) {
	274	mean += delta / (i + 1.0);
	275	S += delta * (cycles[i] - mean);
	276	}
	277	}
	278
e259879e JA	279	/*
	280	* The most common platform clock breakage is returning zero
	281	* indefinitely. Check for that and return failure.
	282	*/
	283	if (!cycles[0] && !cycles[NR_TIME_ITERS - 1])
	284	return 1;
	285
fa80feae	286	S = sqrt(S / (NR_TIME_ITERS - 1.0));
c223da83	287
e51a6629 JA	288	minc = -1ULL;
e51a6629 JA	289	maxc = samples = avg = 0;
fa80feae	290	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	291	double this = cycles[i];
c223da83 JA	292
e51a6629 JA	293	minc = min(cycles[i], minc);
	294	maxc = max(cycles[i], maxc);
	295
03e20d68	296	if ((fmax(this, mean) - fmin(this, mean)) > S)
c223da83 JA	297	continue;
	298	samples++;
	299	avg += this;
	300	}
	301
fa80feae	302	S /= (double) NR_TIME_ITERS;
89db727d	303	mean /= 10.0;
c223da83	304
fa80feae	305	for (i = 0; i < NR_TIME_ITERS; i++)
4b91ee8f JA	306	dprint(FD_TIME, "cycles[%d]=%llu\n", i,
4b91ee8f JA	307	(unsigned long long) cycles[i] / 10);
c223da83	308
d7abad3d	309	avg /= samples;
b0ff22d7	310	avg = (avg + 5) / 10;
e51a6629 JA	311	minc /= 10;
e51a6629 JA	312	maxc /= 10;
4b91ee8f	313	dprint(FD_TIME, "avg: %llu\n", (unsigned long long) avg);
e51a6629 JA	314	dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f\n",
	315	(unsigned long long) minc,
	316	(unsigned long long) maxc, mean, S);
c223da83 JA	317
c223da83 JA	318	cycles_per_usec = avg;
71339117	319	inv_cycles_per_usec = 16777216UL / cycles_per_usec;
b3fa625b	320	max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
71339117	321	dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
96170421	322	#ifdef ARCH_CPU_CLOCK_WRAPS
73df3e07 JA	323	cycles_start = get_cpu_clock();
73df3e07 JA	324	dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
96170421	325	#endif
e259879e	326	return 0;
09a32402 JA	327	}
09a32402 JA	328	#else
e259879e	329	static int calibrate_cpu_clock(void)
09a32402	330	{
919e789d CE	331	#ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
	332	return 0;
	333	#else
e259879e	334	return 1;
09a32402	335	#endif
919e789d CE	336	}
919e789d CE	337	#endif // ARCH_HAVE_CPU_CLOCK
09a32402	338
67bf9823	339	#ifndef CONFIG_TLS_THREAD
5d879392 JA	340	void fio_local_clock_init(int is_thread)
	341	{
	342	struct tv_valid *t;
	343
572cfb3f	344	t = calloc(1, sizeof(*t));
9eb271b9	345	if (pthread_setspecific(tv_tls_key, t)) {
5d879392	346	log_err("fio: can't set TLS key\n");
9eb271b9 JA	347	assert(0);
9eb271b9 JA	348	}
5d879392 JA	349	}
	350
	351	static void kill_tv_tls_key(void *data)
	352	{
	353	free(data);
	354	}
67bf9823 JA	355	#else
	356	void fio_local_clock_init(int is_thread)
	357	{
	358	}
	359	#endif
5d879392	360
09a32402 JA	361	void fio_clock_init(void)
09a32402 JA	362	{
01423eae JA	363	if (fio_clock_source == fio_clock_source_inited)
	364	return;
	365
67bf9823	366	#ifndef CONFIG_TLS_THREAD
5d879392 JA	367	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
5d879392 JA	368	log_err("fio: can't create TLS key\n");
67bf9823	369	#endif
5d879392	370
01423eae	371	fio_clock_source_inited = fio_clock_source;
e259879e JA	372
	373	if (calibrate_cpu_clock())
	374	tsc_reliable = 0;
fa80feae JA	375
	376	/*
	377	* If the arch sets tsc_reliable != 0, then it must be good enough
	378	* to use as THE clock source. For x86 CPUs, this means the TSC
	379	* runs at a constant rate and is synced across CPU cores.
	380	*/
	381	if (tsc_reliable) {
aad918e4	382	if (!fio_clock_source_set && !fio_monotonic_clocktest(0))
fa80feae JA	383	fio_clock_source = CS_CPUCLOCK;
	384	} else if (fio_clock_source == CS_CPUCLOCK)
	385	log_info("fio: clocksource=cpu may not be reliable\n");
c223da83 JA	386	}
c223da83 JA	387
1f440ece	388	uint64_t utime_since(const struct timeval s, const struct timeval e)
be4ecfdf	389	{
39ab7da2	390	long sec, usec;
aa60bc58	391	uint64_t ret;
39ab7da2 JA	392
	393	sec = e->tv_sec - s->tv_sec;
	394	usec = e->tv_usec - s->tv_usec;
	395	if (sec > 0 && usec < 0) {
	396	sec--;
	397	usec += 1000000;
	398	}
	399
	400	/*
	401	* time warp bug on some kernels?
	402	*/
	403	if (sec < 0 \|\| (sec == 0 && usec < 0))
	404	return 0;
	405
	406	ret = sec * 1000000ULL + usec;
	407
	408	return ret;
be4ecfdf JA	409	}
be4ecfdf JA	410
1f440ece	411	uint64_t utime_since_now(const struct timeval *s)
be4ecfdf	412	{
39ab7da2 JA	413	struct timeval t;
	414
	415	fio_gettime(&t, NULL);
	416	return utime_since(s, &t);
be4ecfdf	417	}
783a3eb1	418
1f440ece	419	uint64_t mtime_since(const struct timeval s, const struct timeval e)
783a3eb1	420	{
39ab7da2	421	long sec, usec, ret;
783a3eb1	422
39ab7da2 JA	423	sec = e->tv_sec - s->tv_sec;
	424	usec = e->tv_usec - s->tv_usec;
	425	if (sec > 0 && usec < 0) {
	426	sec--;
	427	usec += 1000000;
783a3eb1 JA	428	}
783a3eb1 JA	429
39ab7da2 JA	430	if (sec < 0 \|\| (sec == 0 && usec < 0))
	431	return 0;
	432
	433	sec *= 1000UL;
	434	usec /= 1000UL;
	435	ret = sec + usec;
	436
	437	return ret;
783a3eb1 JA	438	}
783a3eb1 JA	439
1f440ece	440	uint64_t mtime_since_now(const struct timeval *s)
783a3eb1	441	{
39ab7da2 JA	442	struct timeval t;
39ab7da2 JA	443	void *p = __builtin_return_address(0);
783a3eb1	444
39ab7da2 JA	445	fio_gettime(&t, p);
	446	return mtime_since(s, &t);
	447	}
783a3eb1	448
1f440ece	449	uint64_t time_since_now(const struct timeval *s)
39ab7da2 JA	450	{
39ab7da2 JA	451	return mtime_since_now(s) / 1000;
783a3eb1	452	}
7d11f871	453
67bf9823 JA	454	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
67bf9823 JA	455	defined(CONFIG_SFAA)
7d11f871	456
aad918e4 JA	457	#define CLOCK_ENTRIES_DEBUG 100000
aad918e4 JA	458	#define CLOCK_ENTRIES_TEST 10000
7d11f871 JA	459
7d11f871 JA	460	struct clock_entry {
58002f9a JA	461	uint32_t seq;
58002f9a JA	462	uint32_t cpu;
ba458c2f	463	uint64_t tsc;
7d11f871 JA	464	};
	465
	466	struct clock_thread {
	467	pthread_t thread;
	468	int cpu;
aad918e4	469	int debug;
7d11f871 JA	470	pthread_mutex_t lock;
7d11f871 JA	471	pthread_mutex_t started;
aad918e4	472	unsigned long nr_entries;
58002f9a	473	uint32_t *seq;
7d11f871 JA	474	struct clock_entry *entries;
	475	};
	476
58002f9a	477	static inline uint32_t atomic32_inc_return(uint32_t *seq)
7d11f871 JA	478	{
	479	return 1 + __sync_fetch_and_add(seq, 1);
	480	}
	481
	482	static void clock_thread_fn(void data)
	483	{
	484	struct clock_thread *t = data;
	485	struct clock_entry *c;
	486	os_cpu_mask_t cpu_mask;
58002f9a	487	uint32_t last_seq;
7d11f871 JA	488	int i;
7d11f871 JA	489
c763aea6	490	if (fio_cpuset_init(&cpu_mask)) {
37e20021	491	int __err = errno;
c763aea6 JA	492
	493	log_err("clock cpuset init failed: %s\n", strerror(__err));
	494	goto err_out;
	495	}
	496
7d11f871 JA	497	fio_cpu_set(&cpu_mask, t->cpu);
	498
	499	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
e66d7f90 JA	500	int __err = errno;
	501
	502	log_err("clock setaffinity failed: %s\n", strerror(__err));
c763aea6	503	goto err;
7d11f871 JA	504	}
7d11f871 JA	505
7d11f871	506	pthread_mutex_lock(&t->lock);
b9b3498e	507	pthread_mutex_unlock(&t->started);
7d11f871	508
58002f9a	509	last_seq = 0;
7d11f871	510	c = &t->entries[0];
aad918e4	511	for (i = 0; i < t->nr_entries; i++, c++) {
58002f9a JA	512	uint32_t seq;
58002f9a JA	513	uint64_t tsc;
7d11f871 JA	514
	515	c->cpu = t->cpu;
	516	do {
58002f9a JA	517	seq = atomic32_inc_return(t->seq);
	518	if (seq < last_seq)
	519	break;
7d11f871 JA	520	tsc = get_cpu_clock();
	521	} while (seq != *t->seq);
	522
	523	c->seq = seq;
	524	c->tsc = tsc;
	525	}
	526
aad918e4 JA	527	if (t->debug) {
	528	unsigned long long clocks;
	529
	530	clocks = t->entries[i - 1].tsc - t->entries[0].tsc;
	531	log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu, clocks);
	532	}
58002f9a	533
e259879e JA	534	/*
	535	* The most common platform clock breakage is returning zero
	536	* indefinitely. Check for that and return failure.
	537	*/
58002f9a	538	if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
c763aea6	539	goto err;
e259879e	540
d47d7cb3	541	fio_cpuset_exit(&cpu_mask);
7d11f871	542	return NULL;
c763aea6 JA	543	err:
	544	fio_cpuset_exit(&cpu_mask);
	545	err_out:
	546	return (void *) 1;
7d11f871 JA	547	}
	548
	549	static int clock_cmp(const void p1, const void p2)
	550	{
	551	const struct clock_entry *c1 = p1;
	552	const struct clock_entry *c2 = p2;
	553
b9b3498e JA	554	if (c1->seq == c2->seq)
	555	log_err("cs: bug in atomic sequence!\n");
	556
7d11f871 JA	557	return c1->seq - c2->seq;
	558	}
	559
aad918e4	560	int fio_monotonic_clocktest(int debug)
7d11f871	561	{
8a1db9a1	562	struct clock_thread *cthreads;
7d11f871 JA	563	unsigned int nr_cpus = cpus_online();
7d11f871 JA	564	struct clock_entry *entries;
aad918e4	565	unsigned long nr_entries, tentries, failed = 0;
80da8a8f	566	struct clock_entry prev, this;
58002f9a	567	uint32_t seq = 0;
caa3eb1c	568	unsigned int i;
7d11f871	569
aad918e4 JA	570	if (debug) {
aad918e4 JA	571	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
d5e3f5d8	572
b5b571a3	573	#ifdef FIO_INC_DEBUG
aad918e4	574	fio_debug \|= 1U << FD_TIME;
b5b571a3	575	#endif
aad918e4 JA	576	nr_entries = CLOCK_ENTRIES_DEBUG;
	577	} else
	578	nr_entries = CLOCK_ENTRIES_TEST;
	579
4f1d43c2	580	calibrate_cpu_clock();
aad918e4 JA	581
aad918e4 JA	582	if (debug) {
b5b571a3	583	#ifdef FIO_INC_DEBUG
aad918e4	584	fio_debug &= ~(1U << FD_TIME);
b5b571a3	585	#endif
aad918e4	586	}
4f1d43c2	587
8a1db9a1	588	cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
aad918e4	589	tentries = nr_entries * nr_cpus;
7d11f871 JA	590	entries = malloc(tentries * sizeof(struct clock_entry));
7d11f871 JA	591
aad918e4 JA	592	if (debug)
aad918e4 JA	593	log_info("cs: Testing %u CPUs\n", nr_cpus);
7d11f871 JA	594
7d11f871 JA	595	for (i = 0; i < nr_cpus; i++) {
8a1db9a1	596	struct clock_thread *t = &cthreads[i];
7d11f871 JA	597
7d11f871 JA	598	t->cpu = i;
aad918e4	599	t->debug = debug;
7d11f871	600	t->seq = &seq;
aad918e4 JA	601	t->nr_entries = nr_entries;
aad918e4 JA	602	t->entries = &entries[i * nr_entries];
7d11f871 JA	603	pthread_mutex_init(&t->lock, NULL);
	604	pthread_mutex_init(&t->started, NULL);
	605	pthread_mutex_lock(&t->lock);
6b0110cd JA	606	if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
	607	failed++;
	608	nr_cpus = i;
	609	break;
	610	}
7d11f871 JA	611	}
	612
	613	for (i = 0; i < nr_cpus; i++) {
8a1db9a1	614	struct clock_thread *t = &cthreads[i];
7d11f871 JA	615
	616	pthread_mutex_lock(&t->started);
	617	}
	618
	619	for (i = 0; i < nr_cpus; i++) {
8a1db9a1	620	struct clock_thread *t = &cthreads[i];
7d11f871 JA	621
	622	pthread_mutex_unlock(&t->lock);
	623	}
	624
814917be	625	for (i = 0; i < nr_cpus; i++) {
8a1db9a1	626	struct clock_thread *t = &cthreads[i];
7d11f871 JA	627	void *ret;
	628
	629	pthread_join(t->thread, &ret);
	630	if (ret)
	631	failed++;
	632	}
8a1db9a1	633	free(cthreads);
7d11f871 JA	634
7d11f871 JA	635	if (failed) {
aad918e4 JA	636	if (debug)
aad918e4 JA	637	log_err("Clocksource test: %lu threads failed\n", failed);
7d11f871 JA	638	goto err;
	639	}
	640
	641	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	642
a4aa3dc0 JA	643	/* silence silly gcc */
a4aa3dc0 JA	644	prev = NULL;
7d11f871	645	for (failed = i = 0; i < tentries; i++) {
80da8a8f	646	this = &entries[i];
7d11f871 JA	647
	648	if (!i) {
	649	prev = this;
	650	continue;
	651	}
	652
	653	if (prev->tsc > this->tsc) {
	654	uint64_t diff = prev->tsc - this->tsc;
	655
aad918e4 JA	656	if (!debug) {
	657	failed++;
	658	break;
	659	}
	660
4e0a8fa2 JA	661	log_info("cs: CPU clock mismatch (diff=%llu):\n",
	662	(unsigned long long) diff);
	663	log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", prev->cpu, (unsigned long long) prev->tsc, prev->seq);
	664	log_info("\t CPU%3u: TSC=%llu, SEQ=%u\n", this->cpu, (unsigned long long) this->tsc, this->seq);
7d11f871 JA	665	failed++;
	666	}
	667
	668	prev = this;
	669	}
	670
aad918e4 JA	671	if (debug) {
	672	if (failed)
	673	log_info("cs: Failed: %lu\n", failed);
	674	else
	675	log_info("cs: Pass!\n");
	676	}
7d11f871 JA	677	err:
	678	free(entries);
	679	return !!failed;
	680	}
	681
	682	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
	683
aad918e4	684	int fio_monotonic_clocktest(int debug)
7d11f871	685	{
aad918e4 JA	686	if (debug)
	687	log_info("cs: current platform does not support CPU clocks\n");
	688	return 1;
7d11f871 JA	689	}
	690
	691	#endif