#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;
+
+struct tv_valid {
+ struct timeval last_tv;
+ int last_tv_valid;
+};
+static pthread_key_t tv_tls_key;
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
void fio_gettime(struct timeval *tp, void fio_unused *caller)
#endif
{
+ struct tv_valid *t;
+
#ifdef FIO_DEBUG_TIME
if (!caller)
caller = __builtin_return_address(0);
* 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;
+ t = pthread_getspecific(tv_tls_key);
+ if (t) {
+ if (t->last_tv_valid) {
+ if (tp->tv_sec < t->last_tv.tv_sec)
+ tp->tv_sec = t->last_tv.tv_sec;
+ else if (t->last_tv.tv_sec == tp->tv_sec &&
+ tp->tv_usec < t->last_tv.tv_usec)
+ tp->tv_usec = t->last_tv.tv_usec;
+ }
+ t->last_tv_valid = 1;
+ memcpy(&t->last_tv, tp, sizeof(*tp));
}
- last_tv_valid = 1;
- memcpy(&last_tv, tp, sizeof(*tp));
}
#ifdef ARCH_HAVE_CPU_CLOCK
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);
}
#endif
+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);
+}
+
void fio_clock_init(void)
{
- last_tv_valid = 0;
+ if (fio_clock_source == fio_clock_source_inited)
+ return;
+
+ if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
+ log_err("fio: can't create TLS key\n");
+
+ fio_clock_source_inited = fio_clock_source;
calibrate_cpu_clock();
/*
{
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