X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=gettime.c;h=f40abc7a5ad347a00a57092375c3b728ae1fbd16;hp=cc9dcb7bea7ab28215e09304b3c32688be8d8f25;hb=e66d7f908a2681cbdcd718b8c2c0247d0857c229;hpb=836fcc0fceb233ebcc41ee63b4ea5cae20b678a4

diff --git a/gettime.c b/gettime.c
index cc9dcb7b..f40abc7a 100644
--- a/gettime.c
+++ b/gettime.c
@@ -13,26 +13,30 @@
 #include "hash.h"
 #include "os/os.h"
 
-#ifdef ARCH_HAVE_CPU_CLOCK
+#if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
 static unsigned long cycles_per_usec;
 static unsigned long inv_cycles_per_usec;
+static uint64_t max_cycles_for_mult;
+static unsigned long long cycles_start, cycles_wrap;
 #endif
 int tsc_reliable = 0;
 
 struct tv_valid {
-	struct timeval last_tv;
 	uint64_t last_cycles;
 	int last_tv_valid;
+	int warned;
 };
+#ifdef ARCH_HAVE_CPU_CLOCK
 #ifdef CONFIG_TLS_THREAD
-static struct tv_valid __thread static_tv_valid;
+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;
-enum fio_cs fio_clock_source_inited = CS_INVAL;
+static enum fio_cs fio_clock_source_inited = CS_INVAL;
 
 #ifdef FIO_DEBUG_TIME
 
@@ -64,7 +68,7 @@ static struct gtod_log *find_hash(void *caller)
 	return NULL;
 }
 
-static struct gtod_log *find_log(void *caller)
+static void inc_caller(void *caller)
 {
 	struct gtod_log *log = find_hash(caller);
 
@@ -80,16 +84,13 @@ static struct gtod_log *find_log(void *caller)
 		flist_add_tail(&log->list, &hash[h]);
 	}
 
-	return log;
+	log->calls++;
 }
 
 static void gtod_log_caller(void *caller)
 {
-	if (gtod_inited) {
-		struct gtod_log *log = find_log(caller);
-
-		log->calls++;
-	}
+	if (gtod_inited)
+		inc_caller(caller);
 }
 
 static void fio_exit fio_dump_gtod(void)
@@ -136,16 +137,8 @@ static int fill_clock_gettime(struct timespec *ts)
 }
 #endif
 
-static void *__fio_gettime(struct timeval *tp)
+static void __fio_gettime(struct timeval *tp)
 {
-	struct tv_valid *tv;
-
-#ifdef CONFIG_TLS_THREAD
-	tv = &static_tv_valid;
-#else
-	tv = pthread_getspecific(tv_tls_key);
-#endif
-
 	switch (fio_clock_source) {
 #ifdef CONFIG_GETTIMEOFDAY
 	case CS_GTOD:
@@ -169,15 +162,33 @@ static void *__fio_gettime(struct timeval *tp)
 #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();
-		if (tv && t < tv->last_cycles) {
-			dprint(FD_TIME, "CPU clock going back in time\n");
-			t = tv->last_cycles;
-		} else if (tv)
-			tv->last_cycles = t;
+		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;
+		}
 
-		usecs = (t * inv_cycles_per_usec) / 16777216UL;
+		t -= cycles_start;
+		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;
@@ -187,8 +198,6 @@ static void *__fio_gettime(struct timeval *tp)
 		log_err("fio: invalid clock source %d\n", fio_clock_source);
 		break;
 	}
-
-	return tv;
 }
 
 #ifdef FIO_DEBUG_TIME
@@ -197,39 +206,19 @@ void fio_gettime(struct timeval *tp, void *caller)
 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));
+	if (fio_unlikely(fio_gettime_offload(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 (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));
-	}
+	__fio_gettime(tp);
 }
 
-#ifdef ARCH_HAVE_CPU_CLOCK
+#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;
@@ -262,10 +251,10 @@ static unsigned long get_cycles_per_usec(void)
 
 #define NR_TIME_ITERS	50
 
-static void calibrate_cpu_clock(void)
+static int calibrate_cpu_clock(void)
 {
 	double delta, mean, S;
-	uint64_t avg, cycles[NR_TIME_ITERS];
+	uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
 	int i, samples;
 
 	cycles[0] = get_cycles_per_usec();
@@ -279,12 +268,23 @@ 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;
+	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++;
@@ -295,31 +295,47 @@ 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 + 5) / 10;
-	dprint(FD_TIME, "avg: %lu\n", avg);
-	dprint(FD_TIME, "mean=%f, S=%f\n", mean, S);
+	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);
+	cycles_start = get_cpu_clock();
+	dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
+	return 0;
 }
 #else
-static void calibrate_cpu_clock(void)
+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(sizeof(*t), 1);
-	if (pthread_setspecific(tv_tls_key, 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)
@@ -343,7 +359,9 @@ void fio_clock_init(void)
 #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
@@ -351,13 +369,13 @@ void fio_clock_init(void)
 	 * runs at a constant rate and is synced across CPU cores.
 	 */
 	if (tsc_reliable) {
-		if (!fio_clock_source_set)
+		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(struct timeval *s, struct timeval *e)
+uint64_t utime_since(const struct timeval *s, const struct timeval *e)
 {
 	long sec, usec;
 	uint64_t ret;
@@ -380,7 +398,7 @@ uint64_t utime_since(struct timeval *s, struct timeval *e)
 	return ret;
 }
 
-uint64_t utime_since_now(struct timeval *s)
+uint64_t utime_since_now(const struct timeval *s)
 {
 	struct timeval t;
 
@@ -388,7 +406,7 @@ uint64_t utime_since_now(struct timeval *s)
 	return utime_since(s, &t);
 }
 
-uint64_t mtime_since(struct timeval *s, struct timeval *e)
+uint64_t mtime_since(const struct timeval *s, const struct timeval *e)
 {
 	long sec, usec, ret;
 
@@ -409,7 +427,7 @@ uint64_t mtime_since(struct timeval *s, struct timeval *e)
 	return ret;
 }
 
-uint64_t mtime_since_now(struct timeval *s)
+uint64_t mtime_since_now(const struct timeval *s)
 {
 	struct timeval t;
 	void *p = __builtin_return_address(0);
@@ -418,7 +436,7 @@ uint64_t mtime_since_now(struct timeval *s)
 	return mtime_since(s, &t);
 }
 
-uint64_t time_since_now(struct timeval *s)
+uint64_t time_since_now(const struct timeval *s)
 {
 	return mtime_since_now(s) / 1000;
 }
@@ -426,24 +444,27 @@ uint64_t time_since_now(struct timeval *s)
 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
     defined(CONFIG_SFAA)
 
-#define CLOCK_ENTRIES	100000
+#define CLOCK_ENTRIES_DEBUG	100000
+#define CLOCK_ENTRIES_TEST	10000
 
 struct clock_entry {
-	uint64_t seq;
+	uint32_t seq;
+	uint32_t cpu;
 	uint64_t tsc;
-	uint64_t cpu;
 };
 
 struct clock_thread {
 	pthread_t thread;
 	int cpu;
+	int debug;
 	pthread_mutex_t lock;
 	pthread_mutex_t started;
-	uint64_t *seq;
+	unsigned long nr_entries;
+	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);
 }
@@ -453,26 +474,33 @@ static void *clock_thread_fn(void *data)
 	struct clock_thread *t = data;
 	struct clock_entry *c;
 	os_cpu_mask_t cpu_mask;
+	uint32_t last_seq;
 	int i;
 
 	memset(&cpu_mask, 0, sizeof(cpu_mask));
 	fio_cpu_set(&cpu_mask, t->cpu);
 
 	if (fio_setaffinity(gettid(), cpu_mask) == -1) {
-		log_err("clock setaffinity failed\n");
+		int __err = errno;
+
+		log_err("clock setaffinity failed: %s\n", strerror(__err));
 		return (void *) 1;
 	}
 
 	pthread_mutex_lock(&t->lock);
 	pthread_mutex_unlock(&t->started);
 
+	last_seq = 0;
 	c = &t->entries[0];
-	for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
-		uint64_t seq, tsc;
+	for (i = 0; i < t->nr_entries; i++, c++) {
+		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);
 
@@ -480,7 +508,20 @@ 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);
+	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)
+		return (void *) 1;
+
 	return NULL;
 }
 
@@ -495,70 +536,91 @@ static int clock_cmp(const void *p1, const void *p2)
 	return c1->seq - c2->seq;
 }
 
-int fio_monotonic_clocktest(void)
+int fio_monotonic_clocktest(int debug)
 {
-	struct clock_thread *threads;
+	struct clock_thread *cthreads;
 	unsigned int nr_cpus = cpus_online();
 	struct clock_entry *entries;
-	unsigned long tentries, failed;
-	uint64_t seq = 0;
-	int i;
+	unsigned long nr_entries, tentries, failed = 0;
+	struct clock_entry *prev, *this;
+	uint32_t seq = 0;
+	unsigned int i;
 
-	log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
+	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;
 
-	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;
+	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));
 
-	log_info("cs: Testing %u CPUs\n", nr_cpus);
+	if (debug)
+		log_info("cs: Testing %u CPUs\n", nr_cpus);
 
 	for (i = 0; i < nr_cpus; i++) {
-		struct clock_thread *t = &threads[i];
+		struct clock_thread *t = &cthreads[i];
 
 		t->cpu = i;
+		t->debug = debug;
 		t->seq = &seq;
-		t->entries = &entries[i * CLOCK_ENTRIES];
+		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);
-		pthread_create(&t->thread, NULL, clock_thread_fn, t);
+		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 = &threads[i];
+		struct clock_thread *t = &cthreads[i];
 
 		pthread_mutex_lock(&t->started);
 	}
 
 	for (i = 0; i < nr_cpus; i++) {
-		struct clock_thread *t = &threads[i];
+		struct clock_thread *t = &cthreads[i];
 
 		pthread_mutex_unlock(&t->lock);
 	}
 
-	for (failed = i = 0; i < nr_cpus; i++) {
-		struct clock_thread *t = &threads[i];
+	for (i = 0; i < nr_cpus; i++) {
+		struct clock_thread *t = &cthreads[i];
 		void *ret;
 
 		pthread_join(t->thread, &ret);
 		if (ret)
 			failed++;
 	}
-	free(threads);
+	free(cthreads);
 
 	if (failed) {
-		log_err("Clocksource test: %u threads failed\n", failed);
+		if (debug)
+			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;
@@ -568,20 +630,27 @@ 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);
+			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 (failed)
-		log_info("cs: Failed: %lu\n", failed);
-	else
-		log_info("cs: Pass!\n");
-
+	if (debug) {
+		if (failed)
+			log_info("cs: Failed: %lu\n", failed);
+		else
+			log_info("cs: Pass!\n");
+	}
 err:
 	free(entries);
 	return !!failed;
@@ -589,10 +658,11 @@ err:
 
 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 
-int fio_monotonic_clocktest(void)
+int fio_monotonic_clocktest(int debug)
 {
-	log_info("cs: current platform does not support CPU clocks\n");
-	return 0;
+	if (debug)
+		log_info("cs: current platform does not support CPU clocks\n");
+	return 1;
 }
 
 #endif