[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 last_cycles;
int tsc_reliable = 0;
#endif

struct tv_valid {
	struct timeval last_tv;
	int last_tv_valid;
};
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 */

#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 *t;

#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;
	}

	switch (fio_clock_source) {
	case CS_GTOD:
		gettimeofday(tp, NULL);
		break;
	case CS_CGETTIME: {
		struct timespec ts;

#ifdef FIO_HAVE_CLOCK_MONOTONIC
		if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
#else
		if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
#endif
			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 (t < last_cycles) {
			dprint(FD_TIME, "CPU clock going back in time\n");
			t = last_cycles;
		}

		usecs = t / cycles_per_usec;
		tp->tv_sec = usecs / 1000000;
		tp->tv_usec = usecs % 1000000;
		last_cycles = t;
		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.
	 */
	t = pthread_getspecific(tv_tls_key);
	if (t) {
		if (t->last_tv_valid) {
			if (tp->tv_sec < t->last_tv.tv_sec)
				tp->tv_sec = t->last_tv.tv_sec;
			else if (t->last_tv.tv_sec == tp->tv_sec &&
				 tp->tv_usec < t->last_tv.tv_usec)
				tp->tv_usec = t->last_tv.tv_usec;
		}
		t->last_tv_valid = 1;
		memcpy(&t->last_tv, tp, sizeof(*tp));
	}
}

#ifdef ARCH_HAVE_CPU_CLOCK
static unsigned long get_cycles_per_usec(void)
{
	struct timeval s, e;
	unsigned long long c_s, c_e;

	gettimeofday(&s, NULL);
	c_s = get_cpu_clock();
	do {
		unsigned long long elapsed;

		gettimeofday(&e, NULL);
		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;
}
#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;
c223da83	18	static unsigned long last_cycles;
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;
	25	};
	26	static pthread_key_t tv_tls_key;
02bcaa8c	27
16de1bf9	28	enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
fa80feae	29	int fio_clock_source_set = 0;
01423eae	30	enum fio_cs fio_clock_source_inited = CS_INVAL;
c223da83	31
02bcaa8c JA	32	#ifdef FIO_DEBUG_TIME
	33
	34	#define HASH_BITS 8
	35	#define HASH_SIZE (1 << HASH_BITS)
	36
01743ee1	37	static struct flist_head hash[HASH_SIZE];
02bcaa8c JA	38	static int gtod_inited;
	39
	40	struct gtod_log {
01743ee1	41	struct flist_head list;
02bcaa8c JA	42	void *caller;
	43	unsigned long calls;
	44	};
	45
	46	static struct gtod_log find_hash(void caller)
	47	{
	48	unsigned long h = hash_ptr(caller, HASH_BITS);
01743ee1	49	struct flist_head *entry;
02bcaa8c	50
01743ee1 JA	51	flist_for_each(entry, &hash[h]) {
	52	struct gtod_log *log = flist_entry(entry, struct gtod_log,
	53	list);
02bcaa8c JA	54
	55	if (log->caller == caller)
	56	return log;
	57	}
	58
	59	return NULL;
	60	}
	61
	62	static struct gtod_log find_log(void caller)
	63	{
	64	struct gtod_log *log = find_hash(caller);
	65
	66	if (!log) {
	67	unsigned long h;
	68
	69	log = malloc(sizeof(*log));
01743ee1	70	INIT_FLIST_HEAD(&log->list);
02bcaa8c JA	71	log->caller = caller;
	72	log->calls = 0;
	73
	74	h = hash_ptr(caller, HASH_BITS);
01743ee1	75	flist_add_tail(&log->list, &hash[h]);
02bcaa8c JA	76	}
	77
	78	return log;
	79	}
	80
	81	static void gtod_log_caller(void *caller)
	82	{
	83	if (gtod_inited) {
	84	struct gtod_log *log = find_log(caller);
	85
	86	log->calls++;
	87	}
	88	}
	89
	90	static void fio_exit fio_dump_gtod(void)
	91	{
	92	unsigned long total_calls = 0;
	93	int i;
	94
	95	for (i = 0; i < HASH_SIZE; i++) {
01743ee1	96	struct flist_head *entry;
02bcaa8c JA	97	struct gtod_log *log;
02bcaa8c JA	98
01743ee1 JA	99	flist_for_each(entry, &hash[i]) {
01743ee1 JA	100	log = flist_entry(entry, struct gtod_log, list);
02bcaa8c	101
5ec10eaa JA	102	printf("function %p, calls %lu\n", log->caller,
5ec10eaa JA	103	log->calls);
02bcaa8c JA	104	total_calls += log->calls;
	105	}
	106	}
	107
	108	printf("Total %lu gettimeofday\n", total_calls);
	109	}
	110
	111	static void fio_init gtod_init(void)
	112	{
	113	int i;
	114
	115	for (i = 0; i < HASH_SIZE; i++)
01743ee1	116	INIT_FLIST_HEAD(&hash[i]);
02bcaa8c JA	117
	118	gtod_inited = 1;
	119	}
	120
	121	#endif /* FIO_DEBUG_TIME */
	122
1e97cce9	123	#ifdef FIO_DEBUG_TIME
02bcaa8c	124	void fio_gettime(struct timeval tp, void caller)
1e97cce9 JA	125	#else
	126	void fio_gettime(struct timeval tp, void fio_unused caller)
	127	#endif
02bcaa8c	128	{
5d879392 JA	129	struct tv_valid *t;
5d879392 JA	130
02bcaa8c JA	131	#ifdef FIO_DEBUG_TIME
	132	if (!caller)
	133	caller = __builtin_return_address(0);
	134
	135	gtod_log_caller(caller);
02bcaa8c	136	#endif
be4ecfdf JA	137	if (fio_tv) {
	138	memcpy(tp, fio_tv, sizeof(*tp));
	139	return;
c223da83 JA	140	}
	141
	142	switch (fio_clock_source) {
	143	case CS_GTOD:
02bcaa8c	144	gettimeofday(tp, NULL);
c223da83 JA	145	break;
c223da83 JA	146	case CS_CGETTIME: {
02bcaa8c JA	147	struct timespec ts;
02bcaa8c JA	148
5bd9c78c JA	149	#ifdef FIO_HAVE_CLOCK_MONOTONIC
	150	if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
	151	#else
d481e006	152	if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
5bd9c78c	153	#endif
c223da83 JA	154	log_err("fio: clock_gettime fails\n");
c223da83 JA	155	assert(0);
02bcaa8c JA	156	}
	157
	158	tp->tv_sec = ts.tv_sec;
	159	tp->tv_usec = ts.tv_nsec / 1000;
c223da83 JA	160	break;
	161	}
	162	#ifdef ARCH_HAVE_CPU_CLOCK
	163	case CS_CPUCLOCK: {
	164	unsigned long long usecs, t;
	165
	166	t = get_cpu_clock();
	167	if (t < last_cycles) {
	168	dprint(FD_TIME, "CPU clock going back in time\n");
	169	t = last_cycles;
	170	}
	171
	172	usecs = t / cycles_per_usec;
	173	tp->tv_sec = usecs / 1000000;
	174	tp->tv_usec = usecs % 1000000;
	175	last_cycles = t;
	176	break;
	177	}
	178	#endif
	179	default:
	180	log_err("fio: invalid clock source %d\n", fio_clock_source);
	181	break;
02bcaa8c	182	}
3e488920 JA	183
	184	/*
	185	* If Linux is using the tsc clock on non-synced processors,
	186	* sometimes time can appear to drift backwards. Fix that up.
	187	*/
5d879392 JA	188	t = pthread_getspecific(tv_tls_key);
	189	if (t) {
	190	if (t->last_tv_valid) {
	191	if (tp->tv_sec < t->last_tv.tv_sec)
	192	tp->tv_sec = t->last_tv.tv_sec;
	193	else if (t->last_tv.tv_sec == tp->tv_sec &&
	194	tp->tv_usec < t->last_tv.tv_usec)
	195	tp->tv_usec = t->last_tv.tv_usec;
	196	}
	197	t->last_tv_valid = 1;
	198	memcpy(&t->last_tv, tp, sizeof(*tp));
3e488920	199	}
02bcaa8c	200	}
be4ecfdf	201
09a32402	202	#ifdef ARCH_HAVE_CPU_CLOCK
c223da83 JA	203	static unsigned long get_cycles_per_usec(void)
	204	{
	205	struct timeval s, e;
	206	unsigned long long c_s, c_e;
	207
	208	gettimeofday(&s, NULL);
	209	c_s = get_cpu_clock();
	210	do {
	211	unsigned long long elapsed;
	212
	213	gettimeofday(&e, NULL);
	214	elapsed = utime_since(&s, &e);
486332e5	215	if (elapsed >= 1280) {
c223da83 JA	216	c_e = get_cpu_clock();
	217	break;
	218	}
	219	} while (1);
	220
c011000d	221	return (c_e - c_s + 127) >> 7;
c223da83 JA	222	}
c223da83 JA	223
fa80feae JA	224	#define NR_TIME_ITERS 50
fa80feae JA	225
09a32402	226	static void calibrate_cpu_clock(void)
c223da83 JA	227	{
c223da83 JA	228	double delta, mean, S;
fa80feae	229	unsigned long avg, cycles[NR_TIME_ITERS];
c223da83 JA	230	int i, samples;
c223da83 JA	231
c223da83 JA	232	cycles[0] = get_cycles_per_usec();
c223da83 JA	233	S = delta = mean = 0.0;
fa80feae	234	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	235	cycles[i] = get_cycles_per_usec();
	236	delta = cycles[i] - mean;
	237	if (delta) {
	238	mean += delta / (i + 1.0);
	239	S += delta * (cycles[i] - mean);
	240	}
	241	}
	242
fa80feae	243	S = sqrt(S / (NR_TIME_ITERS - 1.0));
c223da83 JA	244
c223da83 JA	245	samples = avg = 0;
fa80feae	246	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	247	double this = cycles[i];
c223da83 JA	248
03e20d68	249	if ((fmax(this, mean) - fmin(this, mean)) > S)
c223da83 JA	250	continue;
	251	samples++;
	252	avg += this;
	253	}
	254
fa80feae	255	S /= (double) NR_TIME_ITERS;
89db727d	256	mean /= 10.0;
c223da83	257
fa80feae	258	for (i = 0; i < NR_TIME_ITERS; i++)
c223da83 JA	259	dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
c223da83 JA	260
d7abad3d JA	261	avg /= samples;
d7abad3d JA	262	avg = (avg + 9) / 10;
c223da83 JA	263	dprint(FD_TIME, "avg: %lu\n", avg);
	264	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
	265
	266	cycles_per_usec = avg;
09a32402 JA	267	}
	268	#else
	269	static void calibrate_cpu_clock(void)
	270	{
	271	}
	272	#endif
	273
5d879392 JA	274	void fio_local_clock_init(int is_thread)
	275	{
	276	struct tv_valid *t;
	277
	278	t = calloc(sizeof(*t), 1);
	279	if (pthread_setspecific(tv_tls_key, t))
	280	log_err("fio: can't set TLS key\n");
	281	}
	282
	283	static void kill_tv_tls_key(void *data)
	284	{
	285	free(data);
	286	}
	287
09a32402 JA	288	void fio_clock_init(void)
09a32402 JA	289	{
01423eae JA	290	if (fio_clock_source == fio_clock_source_inited)
	291	return;
	292
5d879392 JA	293	if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
	294	log_err("fio: can't create TLS key\n");
	295
01423eae	296	fio_clock_source_inited = fio_clock_source;
09a32402	297	calibrate_cpu_clock();
fa80feae JA	298
	299	/*
	300	* If the arch sets tsc_reliable != 0, then it must be good enough
	301	* to use as THE clock source. For x86 CPUs, this means the TSC
	302	* runs at a constant rate and is synced across CPU cores.
	303	*/
	304	if (tsc_reliable) {
	305	if (!fio_clock_source_set)
	306	fio_clock_source = CS_CPUCLOCK;
	307	} else if (fio_clock_source == CS_CPUCLOCK)
	308	log_info("fio: clocksource=cpu may not be reliable\n");
c223da83 JA	309	}
c223da83 JA	310
39ab7da2	311	unsigned long long utime_since(struct timeval s, struct timeval e)
be4ecfdf	312	{
39ab7da2 JA	313	long sec, usec;
	314	unsigned long long ret;
	315
	316	sec = e->tv_sec - s->tv_sec;
	317	usec = e->tv_usec - s->tv_usec;
	318	if (sec > 0 && usec < 0) {
	319	sec--;
	320	usec += 1000000;
	321	}
	322
	323	/*
	324	* time warp bug on some kernels?
	325	*/
	326	if (sec < 0 \|\| (sec == 0 && usec < 0))
	327	return 0;
	328
	329	ret = sec * 1000000ULL + usec;
	330
	331	return ret;
be4ecfdf JA	332	}
be4ecfdf JA	333
39ab7da2	334	unsigned long long utime_since_now(struct timeval *s)
be4ecfdf	335	{
39ab7da2 JA	336	struct timeval t;
	337
	338	fio_gettime(&t, NULL);
	339	return utime_since(s, &t);
be4ecfdf	340	}
783a3eb1	341
39ab7da2	342	unsigned long mtime_since(struct timeval s, struct timeval e)
783a3eb1	343	{
39ab7da2	344	long sec, usec, ret;
783a3eb1	345
39ab7da2 JA	346	sec = e->tv_sec - s->tv_sec;
	347	usec = e->tv_usec - s->tv_usec;
	348	if (sec > 0 && usec < 0) {
	349	sec--;
	350	usec += 1000000;
783a3eb1 JA	351	}
783a3eb1 JA	352
39ab7da2 JA	353	if (sec < 0 \|\| (sec == 0 && usec < 0))
	354	return 0;
	355
	356	sec *= 1000UL;
	357	usec /= 1000UL;
	358	ret = sec + usec;
	359
	360	return ret;
783a3eb1 JA	361	}
783a3eb1 JA	362
39ab7da2	363	unsigned long mtime_since_now(struct timeval *s)
783a3eb1	364	{
39ab7da2 JA	365	struct timeval t;
39ab7da2 JA	366	void *p = __builtin_return_address(0);
783a3eb1	367
39ab7da2 JA	368	fio_gettime(&t, p);
	369	return mtime_since(s, &t);
	370	}
783a3eb1	371
39ab7da2 JA	372	unsigned long time_since_now(struct timeval *s)
	373	{
	374	return mtime_since_now(s) / 1000;
783a3eb1	375	}
7d11f871 JA	376
	377	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)
	378
	379	#define CLOCK_ENTRIES 100000
	380
	381	struct clock_entry {
	382	unsigned long seq;
	383	unsigned long tsc;
	384	unsigned long cpu;
	385	};
	386
	387	struct clock_thread {
	388	pthread_t thread;
	389	int cpu;
	390	pthread_mutex_t lock;
	391	pthread_mutex_t started;
	392	uint64_t *seq;
	393	struct clock_entry *entries;
	394	};
	395
	396	static inline uint64_t atomic64_inc_return(uint64_t *seq)
	397	{
	398	return 1 + __sync_fetch_and_add(seq, 1);
	399	}
	400
	401	static void clock_thread_fn(void data)
	402	{
	403	struct clock_thread *t = data;
	404	struct clock_entry *c;
	405	os_cpu_mask_t cpu_mask;
	406	int i;
	407
	408	memset(&cpu_mask, 0, sizeof(cpu_mask));
	409	fio_cpu_set(&cpu_mask, t->cpu);
	410
	411	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
	412	log_err("clock setaffinity failed\n");
	413	return (void *) 1;
	414	}
	415
7d11f871	416	pthread_mutex_lock(&t->lock);
b9b3498e	417	pthread_mutex_unlock(&t->started);
7d11f871 JA	418
	419	c = &t->entries[0];
	420	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
	421	uint64_t seq, tsc;
	422
	423	c->cpu = t->cpu;
	424	do {
	425	seq = atomic64_inc_return(t->seq);
	426	tsc = get_cpu_clock();
	427	} while (seq != *t->seq);
	428
	429	c->seq = seq;
	430	c->tsc = tsc;
	431	}
	432
	433	log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
	434	return NULL;
	435	}
	436
	437	static int clock_cmp(const void p1, const void p2)
	438	{
	439	const struct clock_entry *c1 = p1;
	440	const struct clock_entry *c2 = p2;
	441
b9b3498e JA	442	if (c1->seq == c2->seq)
	443	log_err("cs: bug in atomic sequence!\n");
	444
7d11f871 JA	445	return c1->seq - c2->seq;
	446	}
	447
	448	int fio_monotonic_clocktest(void)
	449	{
	450	struct clock_thread *threads;
	451	unsigned int nr_cpus = cpus_online();
	452	struct clock_entry *entries;
	453	unsigned long tentries, failed;
	454	uint64_t seq = 0;
	455	int i;
	456
4f1d43c2 JA	457	fio_debug \|= 1U << FD_TIME;
	458	calibrate_cpu_clock();
	459	fio_debug &= ~(1U << FD_TIME);
	460
7d11f871 JA	461	threads = malloc(nr_cpus * sizeof(struct clock_thread));
	462	tentries = CLOCK_ENTRIES * nr_cpus;
	463	entries = malloc(tentries * sizeof(struct clock_entry));
	464
	465	log_info("cs: Testing %u CPUs\n", nr_cpus);
	466
	467	for (i = 0; i < nr_cpus; i++) {
	468	struct clock_thread *t = &threads[i];
	469
	470	t->cpu = i;
	471	t->seq = &seq;
	472	t->entries = &entries[i * CLOCK_ENTRIES];
	473	pthread_mutex_init(&t->lock, NULL);
	474	pthread_mutex_init(&t->started, NULL);
	475	pthread_mutex_lock(&t->lock);
	476	pthread_create(&t->thread, NULL, clock_thread_fn, t);
	477	}
	478
	479	for (i = 0; i < nr_cpus; i++) {
	480	struct clock_thread *t = &threads[i];
	481
	482	pthread_mutex_lock(&t->started);
	483	}
	484
	485	for (i = 0; i < nr_cpus; i++) {
	486	struct clock_thread *t = &threads[i];
	487
	488	pthread_mutex_unlock(&t->lock);
	489	}
	490
	491	for (failed = i = 0; i < nr_cpus; i++) {
	492	struct clock_thread *t = &threads[i];
	493	void *ret;
	494
	495	pthread_join(t->thread, &ret);
	496	if (ret)
	497	failed++;
	498	}
	499	free(threads);
	500
	501	if (failed) {
	502	log_err("Clocksource test: %u threads failed\n", failed);
	503	goto err;
	504	}
	505
	506	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	507
	508	for (failed = i = 0; i < tentries; i++) {
	509	struct clock_entry prev, this = &entries[i];
	510
	511	if (!i) {
	512	prev = this;
	513	continue;
	514	}
	515
	516	if (prev->tsc > this->tsc) {
	517	uint64_t diff = prev->tsc - this->tsc;
	518
	519	log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
	520	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
	521	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
	522	failed++;
	523	}
	524
525	prev = this;
526	}
527
528	if (failed)
529	log_info("cs: Failed: %lu\n", failed);
530	else
531	log_info("cs: Pass!\n");
532
533	err:
534	free(entries);
535	return !!failed;
536	}
537
538	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
539
540	int fio_monotonic_clocktest(void)
541	{
542	log_info("cs: current platform does not support CPU clocks\n");
543	return 0;
544	}
545
546	#endif