#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 struct timeval *fio_tv;
-int fio_gtod_offload = 0;
-int fio_gtod_cpu = -1;
-static pthread_t gtod_thread;
+static __thread struct timeval last_tv;
+static __thread int last_tv_valid;
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
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) {
}
}
- 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)
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)
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");
}
-void fio_gtod_init(void)
+unsigned long long utime_since(struct timeval *s, struct timeval *e)
{
- fio_tv = smalloc(sizeof(struct timeval));
- assert(fio_tv);
+ 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;
}
-static void fio_gtod_update(void)
+unsigned long long utime_since_now(struct timeval *s)
{
- gettimeofday(fio_tv, NULL);
+ struct timeval t;
+
+ fio_gettime(&t, NULL);
+ return utime_since(s, &t);
}
-static void *gtod_thread_main(void *data)
+unsigned long mtime_since(struct timeval *s, struct timeval *e)
{
- struct fio_mutex *mutex = data;
+ long sec, usec, ret;
- fio_mutex_up(mutex);
+ sec = e->tv_sec - s->tv_sec;
+ usec = e->tv_usec - s->tv_usec;
+ if (sec > 0 && usec < 0) {
+ sec--;
+ usec += 1000000;
+ }
- /*
- * As long as we have jobs around, update the clock. It would be nice
- * to have some way of NOT hammering that CPU with gettimeofday(),
- * but I'm not sure what to use outside of a simple CPU nop to relax
- * it - we don't want to lose precision.
- */
- while (threads) {
- fio_gtod_update();
- nop;
+ 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;
}
-int fio_start_gtod_thread(void)
+static int clock_cmp(const void *p1, const void *p2)
{
- struct fio_mutex *mutex;
- pthread_attr_t attr;
- int ret;
-
- mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
- if (!mutex)
- return 1;
-
- pthread_attr_init(&attr);
- pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
- ret = pthread_create(>od_thread, &attr, gtod_thread_main, NULL);
- pthread_attr_destroy(&attr);
- if (ret) {
- log_err("Can't create gtod thread: %s\n", strerror(ret));
- goto err;
+ 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);
- ret = pthread_detach(gtod_thread);
- if (ret) {
- log_err("Can't detatch gtod thread: %s\n", strerror(ret));
+ if (failed) {
+ log_err("Clocksource test: %u threads failed\n", failed);
goto err;
}
- dprint(FD_MUTEX, "wait on startup_mutex\n");
- fio_mutex_down(mutex);
- dprint(FD_MUTEX, "done waiting on startup_mutex\n");
+ 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:
- fio_mutex_remove(mutex);
- return ret;
+ 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
projects / fio.git / blobdiff