X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=gettime.c;h=1a7af377eb24b1aadb31652884ffb312b4235783;hp=72fda3f8597e335be2fc4677a208f02db9fa62b7;hb=1d69e8acd7d338822bfe9630e7cdceee159c7bd2;hpb=03e20d687566753b90383571e5e152c5142bdffd

diff --git a/gettime.c b/gettime.c
index 72fda3f8..1a7af377 100644
--- a/gettime.c
+++ b/gettime.c
@@ -11,19 +11,19 @@
 #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 struct timeval last_tv;
-static int last_tv_valid;
+static __thread struct timeval last_tv;
+static __thread int last_tv_valid;
 
-static struct timeval *fio_tv;
-int fio_gtod_offload = 0;
-int fio_gtod_cpu = -1;
-
-enum fio_cs fio_clock_source = CS_GTOD;
+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
 
@@ -140,7 +140,11 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
 	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);
 		}
@@ -199,24 +203,26 @@ static unsigned long get_cycles_per_usec(void)
 
 		gettimeofday(&e, NULL);
 		elapsed = utime_since(&s, &e);
-		if (elapsed >= 10) {
+		if (elapsed >= 1280) {
 			c_e = get_cpu_clock();
 			break;
 		}
 	} while (1);
 
-	return c_e - c_s;
+	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[10];
+	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 < 10; i++) {
+	for (i = 0; i < NR_TIME_ITERS; i++) {
 		cycles[i] = get_cycles_per_usec();
 		delta = cycles[i] - mean;
 		if (delta) {
@@ -225,10 +231,10 @@ static void calibrate_cpu_clock(void)
 		}
 	}
 
-	S = sqrt(S / (10 - 1.0));
+	S = sqrt(S / (NR_TIME_ITERS - 1.0));
 
 	samples = avg = 0;
-	for (i = 0; i < 10; i++) {
+	for (i = 0; i < NR_TIME_ITERS; i++) {
 		double this = cycles[i];
 
 		if ((fmax(this, mean) - fmin(this, mean)) > S)
@@ -237,18 +243,18 @@ static void calibrate_cpu_clock(void)
 		avg += this;
 	}
 
-	S /= 10.0;
+	S /= (double) NR_TIME_ITERS;
 	mean /= 10.0;
 
-	for (i = 0; i < 10; i++)
+	for (i = 0; i < NR_TIME_ITERS; i++)
 		dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
 
-	avg /= (samples * 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)
@@ -258,17 +264,258 @@ static void calibrate_cpu_clock(void)
 
 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;
 }
 
-void fio_gtod_init(void)
+static int clock_cmp(const void *p1, const void *p2)
 {
-	fio_tv = smalloc(sizeof(struct timeval));
-	assert(fio_tv);
+	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;
 }
 
-void fio_gtod_update(void)
+#else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
+
+int fio_monotonic_clocktest(void)
 {
-	gettimeofday(fio_tv, NULL);
+	log_info("cs: current platform does not support CPU clocks\n");
+	return 0;
 }
+
+#endif