+ if (fio_clock_source == fio_clock_source_inited)
+ return;
+
+#ifndef CONFIG_TLS_THREAD
+ if (pthread_key_create(&tv_tls_key, kill_tv_tls_key))
+ log_err("fio: can't create TLS key\n");
+#endif
+
+ fio_clock_source_inited = fio_clock_source;
+
+ if (calibrate_cpu_clock())
+ tsc_reliable = 0;
+
+ /*
+ * 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_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(const struct timeval *s, const struct timeval *e)
+{
+ long sec, usec;
+ uint64_t 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;
+}
+
+uint64_t utime_since_now(const struct timeval *s)
+{
+ struct timeval t;
+
+ fio_gettime(&t, NULL);
+ return utime_since(s, &t);
+}
+
+uint64_t mtime_since(const struct timeval *s, const 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;
+}
+
+uint64_t mtime_since_now(const struct timeval *s)
+{
+ struct timeval t;
+ void *p = __builtin_return_address(0);
+
+ fio_gettime(&t, p);
+ return mtime_since(s, &t);
+}
+
+uint64_t time_since_now(const struct timeval *s)
+{
+ return mtime_since_now(s) / 1000;
+}
+
+#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) && \
+ defined(CONFIG_SFAA)
+
+#define CLOCK_ENTRIES_DEBUG 100000
+#define CLOCK_ENTRIES_TEST 10000
+
+struct clock_entry {
+ uint32_t seq;
+ uint32_t cpu;
+ uint64_t tsc;
+};
+
+struct clock_thread {
+ pthread_t thread;
+ int cpu;
+ int debug;
+ pthread_mutex_t lock;
+ pthread_mutex_t started;
+ unsigned long nr_entries;
+ uint32_t *seq;
+ struct clock_entry *entries;
+};
+
+static inline uint32_t atomic32_inc_return(uint32_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;
+ uint32_t last_seq;
+ int i;
+
+ if (fio_cpuset_init(&cpu_mask)) {
+ int __err = errno;
+
+ log_err("clock cpuset init failed: %s\n", strerror(__err));
+ goto err_out;
+ }
+
+ fio_cpu_set(&cpu_mask, t->cpu);
+
+ if (fio_setaffinity(gettid(), cpu_mask) == -1) {
+ int __err = errno;
+
+ log_err("clock setaffinity failed: %s\n", strerror(__err));
+ goto err;
+ }
+
+ pthread_mutex_lock(&t->lock);
+ pthread_mutex_unlock(&t->started);
+
+ last_seq = 0;
+ c = &t->entries[0];
+ for (i = 0; i < t->nr_entries; i++, c++) {
+ uint32_t seq;
+ uint64_t tsc;
+
+ c->cpu = t->cpu;
+ do {
+ seq = atomic32_inc_return(t->seq);
+ if (seq < last_seq)
+ break;
+ tsc = get_cpu_clock();
+ } while (seq != *t->seq);
+
+ c->seq = seq;
+ c->tsc = 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)
+ goto err;
+
+ fio_cpuset_exit(&cpu_mask);
+ return NULL;
+err:
+ fio_cpuset_exit(&cpu_mask);
+err_out:
+ return (void *) 1;
+}
+
+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;