X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=gettime.c;h=1a0f827460c9da6562678733f041be73c2eec484;hp=f5be6bd00d2eff90a6583eaed217bcc6a498b62e;hb=fdc0f3b646e417497849d4398029f780b0e5262f;hpb=4f1d43c22e2a87d0341e607bd1691d57dd2ef2b8

diff --git a/gettime.c b/gettime.c
index f5be6bd0..1a0f8274 100644
--- a/gettime.c
+++ b/gettime.c
@@ -15,11 +15,20 @@
 
 #ifdef ARCH_HAVE_CPU_CLOCK
 static unsigned long cycles_per_usec;
-static unsigned long last_cycles;
+static unsigned long inv_cycles_per_usec;
+#endif
 int tsc_reliable = 0;
+
+struct tv_valid {
+	struct timeval last_tv;
+	uint64_t last_cycles;
+	int last_tv_valid;
+};
+#ifdef CONFIG_TLS_THREAD
+static __thread struct tv_valid static_tv_valid;
+#else
+static pthread_key_t tv_tls_key;
 #endif
-static struct timeval last_tv;
-static int last_tv_valid;
 
 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
 int fio_clock_source_set = 0;
@@ -116,35 +125,38 @@ static void fio_init gtod_init(void)
 
 #endif /* FIO_DEBUG_TIME */
 
-#ifdef FIO_DEBUG_TIME
-void fio_gettime(struct timeval *tp, void *caller)
+#ifdef CONFIG_CLOCK_GETTIME
+static int fill_clock_gettime(struct timespec *ts)
+{
+#ifdef CONFIG_CLOCK_MONOTONIC
+	return clock_gettime(CLOCK_MONOTONIC, ts);
 #else
-void fio_gettime(struct timeval *tp, void fio_unused *caller)
+	return clock_gettime(CLOCK_REALTIME, ts);
+#endif
+}
 #endif
+
+static void *__fio_gettime(struct timeval *tp)
 {
-#ifdef FIO_DEBUG_TIME
-	if (!caller)
-		caller = __builtin_return_address(0);
+	struct tv_valid *tv;
 
-	gtod_log_caller(caller);
+#ifdef CONFIG_TLS_THREAD
+	tv = &static_tv_valid;
+#else
+	tv = pthread_getspecific(tv_tls_key);
 #endif
-	if (fio_tv) {
-		memcpy(tp, fio_tv, sizeof(*tp));
-		return;
-	}
 
 	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;
 
-#ifdef FIO_HAVE_CLOCK_MONOTONIC
-		if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
-#else
-		if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
-#endif
+		if (fill_clock_gettime(&ts) < 0) {
 			log_err("fio: clock_gettime fails\n");
 			assert(0);
 		}
@@ -153,20 +165,21 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
 		tp->tv_usec = ts.tv_nsec / 1000;
 		break;
 		}
+#endif
 #ifdef ARCH_HAVE_CPU_CLOCK
 	case CS_CPUCLOCK: {
-		unsigned long long usecs, t;
+		uint64_t usecs, t;
 
 		t = get_cpu_clock();
-		if (t < last_cycles) {
+		if (tv && t < tv->last_cycles) {
 			dprint(FD_TIME, "CPU clock going back in time\n");
-			t = last_cycles;
-		}
+			t = tv->last_cycles;
+		} else if (tv)
+			tv->last_cycles = t;
 
-		usecs = t / cycles_per_usec;
+		usecs = (t * inv_cycles_per_usec) / 16777216UL;
 		tp->tv_sec = usecs / 1000000;
 		tp->tv_usec = usecs % 1000000;
-		last_cycles = t;
 		break;
 		}
 #endif
@@ -175,33 +188,67 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
 		break;
 	}
 
+	return tv;
+}
+
+#ifdef FIO_DEBUG_TIME
+void fio_gettime(struct timeval *tp, void *caller)
+#else
+void fio_gettime(struct timeval *tp, void fio_unused *caller)
+#endif
+{
+	struct tv_valid *tv;
+
+#ifdef FIO_DEBUG_TIME
+	if (!caller)
+		caller = __builtin_return_address(0);
+
+	gtod_log_caller(caller);
+#endif
+	if (fio_tv) {
+		memcpy(tp, fio_tv, sizeof(*tp));
+		return;
+	}
+
+	tv = __fio_gettime(tp);
+
 	/*
 	 * 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;
+	if (tv) {
+		if (tv->last_tv_valid) {
+			if (tp->tv_sec < tv->last_tv.tv_sec)
+				tp->tv_sec = tv->last_tv.tv_sec;
+			else if (tv->last_tv.tv_sec == tp->tv_sec &&
+				 tp->tv_usec < tv->last_tv.tv_usec)
+				tp->tv_usec = tv->last_tv.tv_usec;
+		}
+		tv->last_tv_valid = 1;
+		memcpy(&tv->last_tv, tp, sizeof(*tp));
 	}
-	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;
+	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);
 
-	gettimeofday(&s, NULL);
 	c_s = get_cpu_clock();
 	do {
-		unsigned long long elapsed;
+		uint64_t elapsed;
+
+		__fio_gettime(&e);
 
-		gettimeofday(&e, NULL);
 		elapsed = utime_since(&s, &e);
 		if (elapsed >= 1280) {
 			c_e = get_cpu_clock();
@@ -209,15 +256,16 @@ static unsigned long get_cycles_per_usec(void)
 		}
 	} while (1);
 
+	fio_clock_source = old_cs;
 	return (c_e - c_s + 127) >> 7;
 }
 
 #define NR_TIME_ITERS	50
 
-static void calibrate_cpu_clock(void)
+static int calibrate_cpu_clock(void)
 {
 	double delta, mean, S;
-	unsigned long avg, cycles[NR_TIME_ITERS];
+	uint64_t avg, cycles[NR_TIME_ITERS];
 	int i, samples;
 
 	cycles[0] = get_cycles_per_usec();
@@ -231,6 +279,13 @@ static void calibrate_cpu_clock(void)
 		}
 	}
 
+	/*
+	 * 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;
@@ -247,17 +302,42 @@ static void calibrate_cpu_clock(void)
 	mean /= 10.0;
 
 	for (i = 0; i < NR_TIME_ITERS; i++)
-		dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
+		dprint(FD_TIME, "cycles[%d]=%llu\n", i,
+					(unsigned long long) cycles[i] / 10);
 
 	avg /= samples;
-	avg = (avg + 9) / 10;
-	dprint(FD_TIME, "avg: %lu\n", avg);
+	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)
+{
+	return 1;
+}
+#endif
+
+#ifndef CONFIG_TLS_THREAD
+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);
 }
 #else
-static void calibrate_cpu_clock(void)
+void fio_local_clock_init(int is_thread)
 {
 }
 #endif
@@ -267,9 +347,15 @@ void fio_clock_init(void)
 	if (fio_clock_source == fio_clock_source_inited)
 		return;
 
-	last_tv_valid = 0;
+#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;
-	calibrate_cpu_clock();
+
+	if (calibrate_cpu_clock())
+		tsc_reliable = 0;
 
 	/*
 	 * If the arch sets tsc_reliable != 0, then it must be good enough
@@ -283,10 +369,10 @@ void fio_clock_init(void)
 		log_info("fio: clocksource=cpu may not be reliable\n");
 }
 
-unsigned long long utime_since(struct timeval *s, struct timeval *e)
+uint64_t utime_since(struct timeval *s, struct timeval *e)
 {
 	long sec, usec;
-	unsigned long long ret;
+	uint64_t ret;
 
 	sec = e->tv_sec - s->tv_sec;
 	usec = e->tv_usec - s->tv_usec;
@@ -306,7 +392,7 @@ unsigned long long utime_since(struct timeval *s, struct timeval *e)
 	return ret;
 }
 
-unsigned long long utime_since_now(struct timeval *s)
+uint64_t utime_since_now(struct timeval *s)
 {
 	struct timeval t;
 
@@ -314,7 +400,7 @@ unsigned long long utime_since_now(struct timeval *s)
 	return utime_since(s, &t);
 }
 
-unsigned long mtime_since(struct timeval *s, struct timeval *e)
+uint64_t mtime_since(struct timeval *s, struct timeval *e)
 {
 	long sec, usec, ret;
 
@@ -335,7 +421,7 @@ unsigned long mtime_since(struct timeval *s, struct timeval *e)
 	return ret;
 }
 
-unsigned long mtime_since_now(struct timeval *s)
+uint64_t mtime_since_now(struct timeval *s)
 {
 	struct timeval t;
 	void *p = __builtin_return_address(0);
@@ -344,19 +430,20 @@ unsigned long mtime_since_now(struct timeval *s)
 	return mtime_since(s, &t);
 }
 
-unsigned long time_since_now(struct timeval *s)
+uint64_t time_since_now(struct timeval *s)
 {
 	return mtime_since_now(s) / 1000;
 }
 
-#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)
+#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
+    defined(CONFIG_SFAA)
 
 #define CLOCK_ENTRIES	100000
 
 struct clock_entry {
-	unsigned long seq;
-	unsigned long tsc;
-	unsigned long cpu;
+	uint32_t seq;
+	uint32_t cpu;
+	uint64_t tsc;
 };
 
 struct clock_thread {
@@ -364,11 +451,11 @@ struct clock_thread {
 	int cpu;
 	pthread_mutex_t lock;
 	pthread_mutex_t started;
-	uint64_t *seq;
+	uint32_t *seq;
 	struct clock_entry *entries;
 };
 
-static inline uint64_t atomic64_inc_return(uint64_t *seq)
+static inline uint32_t atomic32_inc_return(uint32_t *seq)
 {
 	return 1 + __sync_fetch_and_add(seq, 1);
 }
@@ -378,6 +465,7 @@ 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));
@@ -391,13 +479,17 @@ static void *clock_thread_fn(void *data)
 	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++) {
-		uint64_t seq, tsc;
+		uint32_t seq;
+		uint64_t tsc;
 
 		c->cpu = t->cpu;
 		do {
-			seq = atomic64_inc_return(t->seq);
+			seq = atomic32_inc_return(t->seq);
+			if (seq < last_seq)
+				break;
 			tsc = get_cpu_clock();
 		} while (seq != *t->seq);
 
@@ -405,7 +497,16 @@ static void *clock_thread_fn(void *data)
 		c->tsc = tsc;
 	}
 
-	log_info("cs: cpu%3d: %lu clocks seen\n", t->cpu, t->entries[CLOCK_ENTRIES - 1].tsc - t->entries[0].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;
 }
 
@@ -426,9 +527,12 @@ int fio_monotonic_clocktest(void)
 	unsigned int nr_cpus = cpus_online();
 	struct clock_entry *entries;
 	unsigned long tentries, failed;
-	uint64_t seq = 0;
+	struct clock_entry *prev, *this;
+	uint32_t seq = 0;
 	int i;
 
+	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
+
 	fio_debug |= 1U << FD_TIME;
 	calibrate_cpu_clock();
 	fio_debug &= ~(1U << FD_TIME);
@@ -474,14 +578,14 @@ int fio_monotonic_clocktest(void)
 	free(threads);
 
 	if (failed) {
-		log_err("Clocksource test: %u threads failed\n", 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++) {
-		struct clock_entry *prev, *this = &entries[i];
+		this = &entries[i];
 
 		if (!i) {
 			prev = this;
@@ -491,9 +595,10 @@ int fio_monotonic_clocktest(void)
 		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);
+			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++;
 		}