[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
static __thread struct timeval last_tv;
static __thread int last_tv_valid;

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
{
#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.
	 */
	if (last_tv_valid) {
		if (tp->tv_sec < last_tv.tv_sec)
			tp->tv_sec = last_tv.tv_sec;
		else if (last_tv.tv_sec == tp->tv_sec &&
			 tp->tv_usec < last_tv.tv_usec)
			tp->tv_usec = last_tv.tv_usec;
	}
	last_tv_valid = 1;
	memcpy(&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_clock_init(void)
{
	if (fio_clock_source == fio_clock_source_inited)
		return;

	last_tv_valid = 0;
	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
1d69e8ac JA	21	static __thread struct timeval last_tv;
1d69e8ac JA	22	static __thread int last_tv_valid;
02bcaa8c	23
16de1bf9	24	enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
fa80feae	25	int fio_clock_source_set = 0;
01423eae	26	enum fio_cs fio_clock_source_inited = CS_INVAL;
c223da83	27
02bcaa8c JA	28	#ifdef FIO_DEBUG_TIME
	29
	30	#define HASH_BITS 8
	31	#define HASH_SIZE (1 << HASH_BITS)
	32
01743ee1	33	static struct flist_head hash[HASH_SIZE];
02bcaa8c JA	34	static int gtod_inited;
	35
	36	struct gtod_log {
01743ee1	37	struct flist_head list;
02bcaa8c JA	38	void *caller;
	39	unsigned long calls;
	40	};
	41
	42	static struct gtod_log find_hash(void caller)
	43	{
	44	unsigned long h = hash_ptr(caller, HASH_BITS);
01743ee1	45	struct flist_head *entry;
02bcaa8c	46
01743ee1 JA	47	flist_for_each(entry, &hash[h]) {
	48	struct gtod_log *log = flist_entry(entry, struct gtod_log,
	49	list);
02bcaa8c JA	50
	51	if (log->caller == caller)
	52	return log;
	53	}
	54
	55	return NULL;
	56	}
	57
	58	static struct gtod_log find_log(void caller)
	59	{
	60	struct gtod_log *log = find_hash(caller);
	61
	62	if (!log) {
	63	unsigned long h;
	64
	65	log = malloc(sizeof(*log));
01743ee1	66	INIT_FLIST_HEAD(&log->list);
02bcaa8c JA	67	log->caller = caller;
	68	log->calls = 0;
	69
	70	h = hash_ptr(caller, HASH_BITS);
01743ee1	71	flist_add_tail(&log->list, &hash[h]);
02bcaa8c JA	72	}
	73
	74	return log;
	75	}
	76
	77	static void gtod_log_caller(void *caller)
	78	{
	79	if (gtod_inited) {
	80	struct gtod_log *log = find_log(caller);
	81
	82	log->calls++;
	83	}
	84	}
	85
	86	static void fio_exit fio_dump_gtod(void)
	87	{
	88	unsigned long total_calls = 0;
	89	int i;
	90
	91	for (i = 0; i < HASH_SIZE; i++) {
01743ee1	92	struct flist_head *entry;
02bcaa8c JA	93	struct gtod_log *log;
02bcaa8c JA	94
01743ee1 JA	95	flist_for_each(entry, &hash[i]) {
01743ee1 JA	96	log = flist_entry(entry, struct gtod_log, list);
02bcaa8c	97
5ec10eaa JA	98	printf("function %p, calls %lu\n", log->caller,
5ec10eaa JA	99	log->calls);
02bcaa8c JA	100	total_calls += log->calls;
	101	}
	102	}
	103
	104	printf("Total %lu gettimeofday\n", total_calls);
	105	}
	106
	107	static void fio_init gtod_init(void)
	108	{
	109	int i;
	110
	111	for (i = 0; i < HASH_SIZE; i++)
01743ee1	112	INIT_FLIST_HEAD(&hash[i]);
02bcaa8c JA	113
	114	gtod_inited = 1;
	115	}
	116
	117	#endif /* FIO_DEBUG_TIME */
	118
1e97cce9	119	#ifdef FIO_DEBUG_TIME
02bcaa8c	120	void fio_gettime(struct timeval tp, void caller)
1e97cce9 JA	121	#else
	122	void fio_gettime(struct timeval tp, void fio_unused caller)
	123	#endif
02bcaa8c JA	124	{
	125	#ifdef FIO_DEBUG_TIME
	126	if (!caller)
	127	caller = __builtin_return_address(0);
	128
	129	gtod_log_caller(caller);
02bcaa8c	130	#endif
be4ecfdf JA	131	if (fio_tv) {
	132	memcpy(tp, fio_tv, sizeof(*tp));
	133	return;
c223da83 JA	134	}
	135
	136	switch (fio_clock_source) {
	137	case CS_GTOD:
02bcaa8c	138	gettimeofday(tp, NULL);
c223da83 JA	139	break;
c223da83 JA	140	case CS_CGETTIME: {
02bcaa8c JA	141	struct timespec ts;
02bcaa8c JA	142
5bd9c78c JA	143	#ifdef FIO_HAVE_CLOCK_MONOTONIC
	144	if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
	145	#else
d481e006	146	if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
5bd9c78c	147	#endif
c223da83 JA	148	log_err("fio: clock_gettime fails\n");
c223da83 JA	149	assert(0);
02bcaa8c JA	150	}
	151
	152	tp->tv_sec = ts.tv_sec;
	153	tp->tv_usec = ts.tv_nsec / 1000;
c223da83 JA	154	break;
	155	}
	156	#ifdef ARCH_HAVE_CPU_CLOCK
	157	case CS_CPUCLOCK: {
	158	unsigned long long usecs, t;
	159
	160	t = get_cpu_clock();
	161	if (t < last_cycles) {
	162	dprint(FD_TIME, "CPU clock going back in time\n");
	163	t = last_cycles;
	164	}
	165
	166	usecs = t / cycles_per_usec;
	167	tp->tv_sec = usecs / 1000000;
	168	tp->tv_usec = usecs % 1000000;
	169	last_cycles = t;
	170	break;
	171	}
	172	#endif
	173	default:
	174	log_err("fio: invalid clock source %d\n", fio_clock_source);
	175	break;
02bcaa8c	176	}
3e488920 JA	177
	178	/*
	179	* If Linux is using the tsc clock on non-synced processors,
	180	* sometimes time can appear to drift backwards. Fix that up.
	181	*/
	182	if (last_tv_valid) {
	183	if (tp->tv_sec < last_tv.tv_sec)
	184	tp->tv_sec = last_tv.tv_sec;
	185	else if (last_tv.tv_sec == tp->tv_sec &&
	186	tp->tv_usec < last_tv.tv_usec)
	187	tp->tv_usec = last_tv.tv_usec;
	188	}
	189	last_tv_valid = 1;
	190	memcpy(&last_tv, tp, sizeof(*tp));
02bcaa8c	191	}
be4ecfdf	192
09a32402	193	#ifdef ARCH_HAVE_CPU_CLOCK
c223da83 JA	194	static unsigned long get_cycles_per_usec(void)
	195	{
	196	struct timeval s, e;
	197	unsigned long long c_s, c_e;
	198
	199	gettimeofday(&s, NULL);
	200	c_s = get_cpu_clock();
	201	do {
	202	unsigned long long elapsed;
	203
	204	gettimeofday(&e, NULL);
	205	elapsed = utime_since(&s, &e);
486332e5	206	if (elapsed >= 1280) {
c223da83 JA	207	c_e = get_cpu_clock();
	208	break;
	209	}
	210	} while (1);
	211
c011000d	212	return (c_e - c_s + 127) >> 7;
c223da83 JA	213	}
c223da83 JA	214
fa80feae JA	215	#define NR_TIME_ITERS 50
fa80feae JA	216
09a32402	217	static void calibrate_cpu_clock(void)
c223da83 JA	218	{
c223da83 JA	219	double delta, mean, S;
fa80feae	220	unsigned long avg, cycles[NR_TIME_ITERS];
c223da83 JA	221	int i, samples;
c223da83 JA	222
c223da83 JA	223	cycles[0] = get_cycles_per_usec();
c223da83 JA	224	S = delta = mean = 0.0;
fa80feae	225	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	226	cycles[i] = get_cycles_per_usec();
	227	delta = cycles[i] - mean;
	228	if (delta) {
	229	mean += delta / (i + 1.0);
	230	S += delta * (cycles[i] - mean);
	231	}
	232	}
	233
fa80feae	234	S = sqrt(S / (NR_TIME_ITERS - 1.0));
c223da83 JA	235
c223da83 JA	236	samples = avg = 0;
fa80feae	237	for (i = 0; i < NR_TIME_ITERS; i++) {
c223da83 JA	238	double this = cycles[i];
c223da83 JA	239
03e20d68	240	if ((fmax(this, mean) - fmin(this, mean)) > S)
c223da83 JA	241	continue;
	242	samples++;
	243	avg += this;
	244	}
	245
fa80feae	246	S /= (double) NR_TIME_ITERS;
89db727d	247	mean /= 10.0;
c223da83	248
fa80feae	249	for (i = 0; i < NR_TIME_ITERS; i++)
c223da83 JA	250	dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
c223da83 JA	251
d7abad3d JA	252	avg /= samples;
d7abad3d JA	253	avg = (avg + 9) / 10;
c223da83 JA	254	dprint(FD_TIME, "avg: %lu\n", avg);
	255	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
	256
	257	cycles_per_usec = avg;
09a32402 JA	258	}
	259	#else
	260	static void calibrate_cpu_clock(void)
	261	{
	262	}
	263	#endif
	264
	265	void fio_clock_init(void)
	266	{
01423eae JA	267	if (fio_clock_source == fio_clock_source_inited)
	268	return;
	269
09a32402	270	last_tv_valid = 0;
01423eae	271	fio_clock_source_inited = fio_clock_source;
09a32402	272	calibrate_cpu_clock();
fa80feae JA	273
	274	/*
	275	* If the arch sets tsc_reliable != 0, then it must be good enough
	276	* to use as THE clock source. For x86 CPUs, this means the TSC
	277	* runs at a constant rate and is synced across CPU cores.
	278	*/
	279	if (tsc_reliable) {
	280	if (!fio_clock_source_set)
	281	fio_clock_source = CS_CPUCLOCK;
	282	} else if (fio_clock_source == CS_CPUCLOCK)
	283	log_info("fio: clocksource=cpu may not be reliable\n");
c223da83 JA	284	}
c223da83 JA	285
39ab7da2	286	unsigned long long utime_since(struct timeval s, struct timeval e)
be4ecfdf	287	{
39ab7da2 JA	288	long sec, usec;
	289	unsigned long long ret;
	290
	291	sec = e->tv_sec - s->tv_sec;
	292	usec = e->tv_usec - s->tv_usec;
	293	if (sec > 0 && usec < 0) {
	294	sec--;
	295	usec += 1000000;
	296	}
	297
	298	/*
	299	* time warp bug on some kernels?
	300	*/
	301	if (sec < 0 \|\| (sec == 0 && usec < 0))
	302	return 0;
	303
	304	ret = sec * 1000000ULL + usec;
	305
	306	return ret;
be4ecfdf JA	307	}
be4ecfdf JA	308
39ab7da2	309	unsigned long long utime_since_now(struct timeval *s)
be4ecfdf	310	{
39ab7da2 JA	311	struct timeval t;
	312
	313	fio_gettime(&t, NULL);
	314	return utime_since(s, &t);
be4ecfdf	315	}
783a3eb1	316
39ab7da2	317	unsigned long mtime_since(struct timeval s, struct timeval e)
783a3eb1	318	{
39ab7da2	319	long sec, usec, ret;
783a3eb1	320
39ab7da2 JA	321	sec = e->tv_sec - s->tv_sec;
	322	usec = e->tv_usec - s->tv_usec;
	323	if (sec > 0 && usec < 0) {
	324	sec--;
	325	usec += 1000000;
783a3eb1 JA	326	}
783a3eb1 JA	327
39ab7da2 JA	328	if (sec < 0 \|\| (sec == 0 && usec < 0))
	329	return 0;
	330
	331	sec *= 1000UL;
	332	usec /= 1000UL;
	333	ret = sec + usec;
	334
	335	return ret;
783a3eb1 JA	336	}
783a3eb1 JA	337
39ab7da2	338	unsigned long mtime_since_now(struct timeval *s)
783a3eb1	339	{
39ab7da2 JA	340	struct timeval t;
39ab7da2 JA	341	void *p = __builtin_return_address(0);
783a3eb1	342
39ab7da2 JA	343	fio_gettime(&t, p);
	344	return mtime_since(s, &t);
	345	}
783a3eb1	346
39ab7da2 JA	347	unsigned long time_since_now(struct timeval *s)
	348	{
	349	return mtime_since_now(s) / 1000;
783a3eb1	350	}
7d11f871 JA	351
	352	#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)
	353
	354	#define CLOCK_ENTRIES 100000
	355
	356	struct clock_entry {
	357	unsigned long seq;
	358	unsigned long tsc;
	359	unsigned long cpu;
	360	};
	361
	362	struct clock_thread {
	363	pthread_t thread;
	364	int cpu;
	365	pthread_mutex_t lock;
	366	pthread_mutex_t started;
	367	uint64_t *seq;
	368	struct clock_entry *entries;
	369	};
	370
	371	static inline uint64_t atomic64_inc_return(uint64_t *seq)
	372	{
	373	return 1 + __sync_fetch_and_add(seq, 1);
	374	}
	375
	376	static void clock_thread_fn(void data)
	377	{
	378	struct clock_thread *t = data;
	379	struct clock_entry *c;
	380	os_cpu_mask_t cpu_mask;
	381	int i;
	382
	383	memset(&cpu_mask, 0, sizeof(cpu_mask));
	384	fio_cpu_set(&cpu_mask, t->cpu);
	385
	386	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
	387	log_err("clock setaffinity failed\n");
	388	return (void *) 1;
	389	}
	390
7d11f871	391	pthread_mutex_lock(&t->lock);
b9b3498e	392	pthread_mutex_unlock(&t->started);
7d11f871 JA	393
	394	c = &t->entries[0];
	395	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
	396	uint64_t seq, tsc;
	397
	398	c->cpu = t->cpu;
	399	do {
	400	seq = atomic64_inc_return(t->seq);
	401	tsc = get_cpu_clock();
	402	} while (seq != *t->seq);
	403
	404	c->seq = seq;
	405	c->tsc = tsc;
	406	}
	407
	408	log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].tsc);
	409	return NULL;
	410	}
	411
	412	static int clock_cmp(const void p1, const void p2)
	413	{
	414	const struct clock_entry *c1 = p1;
	415	const struct clock_entry *c2 = p2;
	416
b9b3498e JA	417	if (c1->seq == c2->seq)
	418	log_err("cs: bug in atomic sequence!\n");
	419
7d11f871 JA	420	return c1->seq - c2->seq;
	421	}
	422
	423	int fio_monotonic_clocktest(void)
	424	{
	425	struct clock_thread *threads;
	426	unsigned int nr_cpus = cpus_online();
	427	struct clock_entry *entries;
	428	unsigned long tentries, failed;
	429	uint64_t seq = 0;
	430	int i;
	431
4f1d43c2 JA	432	fio_debug \|= 1U << FD_TIME;
	433	calibrate_cpu_clock();
	434	fio_debug &= ~(1U << FD_TIME);
	435
7d11f871 JA	436	threads = malloc(nr_cpus * sizeof(struct clock_thread));
	437	tentries = CLOCK_ENTRIES * nr_cpus;
	438	entries = malloc(tentries * sizeof(struct clock_entry));
	439
	440	log_info("cs: Testing %u CPUs\n", nr_cpus);
	441
	442	for (i = 0; i < nr_cpus; i++) {
	443	struct clock_thread *t = &threads[i];
	444
	445	t->cpu = i;
	446	t->seq = &seq;
	447	t->entries = &entries[i * CLOCK_ENTRIES];
	448	pthread_mutex_init(&t->lock, NULL);
	449	pthread_mutex_init(&t->started, NULL);
	450	pthread_mutex_lock(&t->lock);
	451	pthread_create(&t->thread, NULL, clock_thread_fn, t);
	452	}
	453
	454	for (i = 0; i < nr_cpus; i++) {
	455	struct clock_thread *t = &threads[i];
	456
	457	pthread_mutex_lock(&t->started);
	458	}
	459
	460	for (i = 0; i < nr_cpus; i++) {
	461	struct clock_thread *t = &threads[i];
	462
	463	pthread_mutex_unlock(&t->lock);
	464	}
	465
	466	for (failed = i = 0; i < nr_cpus; i++) {
	467	struct clock_thread *t = &threads[i];
	468	void *ret;
	469
	470	pthread_join(t->thread, &ret);
	471	if (ret)
	472	failed++;
	473	}
	474	free(threads);
	475
	476	if (failed) {
	477	log_err("Clocksource test: %u threads failed\n", failed);
	478	goto err;
	479	}
	480
	481	qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
	482
	483	for (failed = i = 0; i < tentries; i++) {
	484	struct clock_entry prev, this = &entries[i];
	485
	486	if (!i) {
	487	prev = this;
	488	continue;
	489	}
	490
	491	if (prev->tsc > this->tsc) {
	492	uint64_t diff = prev->tsc - this->tsc;
	493
	494	log_info("cs: CPU clock mismatch (diff=%lu):\n", diff);
	495	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", prev->cpu, prev->tsc, prev->seq);
	496	log_info("\t CPU%3lu: TSC=%lu, SEQ=%lu\n", this->cpu, this->tsc, this->seq);
	497	failed++;
	498	}
	499
500	prev = this;
501	}
502
503	if (failed)
504	log_info("cs: Failed: %lu\n", failed);
505	else
506	log_info("cs: Pass!\n");
507
508	err:
509	free(entries);
510	return !!failed;
511	}
512
513	#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
514
515	int fio_monotonic_clocktest(void)
516	{
517	log_info("cs: current platform does not support CPU clocks\n");
518	return 0;
519	}
520
521	#endif