#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;
#endif
int tsc_reliable = 0;
struct tv_valid {
- struct timeval last_tv;
uint64_t last_cycles;
int last_tv_valid;
+ int warned;
};
#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
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
}
#endif
-static void *__fio_gettime(struct timeval *tp)
+static void __fio_gettime(struct timeval *tp)
{
struct tv_valid *tv;
uint64_t usecs, t;
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 < tv->last_cycles && tv->last_tv_valid &&
+ !tv->warned) {
+ log_err("fio: CPU clock going back in time\n");
+ tv->warned = 1;
+ }
+ 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
usecs = (t * inv_cycles_per_usec) / 16777216UL;
+#endif
tp->tv_sec = usecs / 1000000;
tp->tv_usec = usecs % 1000000;
break;
log_err("fio: invalid clock source %d\n", fio_clock_source);
break;
}
-
- return tv;
}
#ifdef FIO_DEBUG_TIME
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;
#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];
}
}
+ /*
+ * 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;
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, "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 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)
#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
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;
return ret;
}
-uint64_t utime_since_now(struct timeval *s)
+uint64_t utime_since_now(const struct timeval *s)
{
struct timeval t;
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;
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);
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;
}
#define CLOCK_ENTRIES 100000
struct clock_entry {
- uint64_t seq;
+ uint32_t seq;
+ uint32_t cpu;
uint64_t tsc;
- uint64_t cpu;
};
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);
}
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));
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);
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;
}
int fio_monotonic_clocktest(void)
{
- 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 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");
+#ifdef FIO_INC_DEBUG
fio_debug |= 1U << FD_TIME;
+#endif
calibrate_cpu_clock();
+#ifdef FIO_INC_DEBUG
fio_debug &= ~(1U << FD_TIME);
+#endif
- threads = malloc(nr_cpus * sizeof(struct clock_thread));
+ cthreads = 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];
+ struct clock_thread *t = &cthreads[i];
t->cpu = i;
t->seq = &seq;
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);
+ 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;
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++;
}
projects / fio.git / blobdiff