genfio: Splitting gen_template in 2 parts
[fio.git] / gettime.c
index 2037354..1a0f827 100644 (file)
--- a/gettime.c
+++ b/gettime.c
 #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;
+static unsigned long inv_cycles_per_usec;
 #endif
-static struct timeval last_tv;
-static int last_tv_valid;
+int tsc_reliable = 0;
 
-static struct timeval *fio_tv;
-int fio_gtod_offload = 0;
-int fio_gtod_cpu = -1;
-static pthread_t gtod_thread;
+struct tv_valid {
+       struct timeval last_tv;
+       uint64_t last_cycles;
+       int last_tv_valid;
+};
+#ifdef CONFIG_TLS_THREAD
+static __thread struct tv_valid static_tv_valid;
+#else
+static pthread_key_t tv_tls_key;
+#endif
 
-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
 
@@ -117,35 +125,38 @@ static void fio_init gtod_init(void)
 
 #endif /* FIO_DEBUG_TIME */
 
-#ifdef FIO_DEBUG_TIME
-void fio_gettime(struct timeval *tp, void *caller)
+#ifdef CONFIG_CLOCK_GETTIME
+static int fill_clock_gettime(struct timespec *ts)
+{
+#ifdef CONFIG_CLOCK_MONOTONIC
+       return clock_gettime(CLOCK_MONOTONIC, ts);
 #else
-void fio_gettime(struct timeval *tp, void fio_unused *caller)
+       return clock_gettime(CLOCK_REALTIME, ts);
 #endif
+}
+#endif
+
+static void *__fio_gettime(struct timeval *tp)
 {
-#ifdef FIO_DEBUG_TIME
-       if (!caller)
-               caller = __builtin_return_address(0);
+       struct tv_valid *tv;
 
-       gtod_log_caller(caller);
+#ifdef CONFIG_TLS_THREAD
+       tv = &static_tv_valid;
+#else
+       tv = pthread_getspecific(tv_tls_key);
 #endif
-       if (fio_tv) {
-               memcpy(tp, fio_tv, sizeof(*tp));
-               return;
-       }
 
        switch (fio_clock_source) {
+#ifdef CONFIG_GETTIMEOFDAY
        case CS_GTOD:
                gettimeofday(tp, NULL);
                break;
+#endif
+#ifdef CONFIG_CLOCK_GETTIME
        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
+               if (fill_clock_gettime(&ts) < 0) {
                        log_err("fio: clock_gettime fails\n");
                        assert(0);
                }
@@ -154,20 +165,21 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
                tp->tv_usec = ts.tv_nsec / 1000;
                break;
                }
+#endif
 #ifdef ARCH_HAVE_CPU_CLOCK
        case CS_CPUCLOCK: {
-               unsigned long long usecs, t;
+               uint64_t usecs, t;
 
                t = get_cpu_clock();
-               if (t < last_cycles) {
+               if (tv && t < tv->last_cycles) {
                        dprint(FD_TIME, "CPU clock going back in time\n");
-                       t = last_cycles;
-               }
+                       t = tv->last_cycles;
+               } else if (tv)
+                       tv->last_cycles = t;
 
-               usecs = t / cycles_per_usec;
+               usecs = (t * inv_cycles_per_usec) / 16777216UL;
                tp->tv_sec = usecs / 1000000;
                tp->tv_usec = usecs % 1000000;
-               last_cycles = t;
                break;
                }
 #endif
@@ -176,52 +188,89 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
                break;
        }
 
+       return tv;
+}
+
+#ifdef FIO_DEBUG_TIME
+void fio_gettime(struct timeval *tp, void *caller)
+#else
+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));
+               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 (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;
+       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));
        }
-       last_tv_valid = 1;
-       memcpy(&last_tv, tp, sizeof(*tp));
 }
 
 #ifdef ARCH_HAVE_CPU_CLOCK
 static unsigned long get_cycles_per_usec(void)
 {
        struct timeval s, e;
-       unsigned long long c_s, c_e;
+       uint64_t c_s, c_e;
+       enum fio_cs old_cs = fio_clock_source;
+
+#ifdef CONFIG_CLOCK_GETTIME
+       fio_clock_source = CS_CGETTIME;
+#else
+       fio_clock_source = CS_GTOD;
+#endif
+       __fio_gettime(&s);
 
-       gettimeofday(&s, NULL);
        c_s = get_cpu_clock();
        do {
-               unsigned long long elapsed;
+               uint64_t elapsed;
+
+               __fio_gettime(&e);
 
-               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;
+       fio_clock_source = old_cs;
+       return (c_e - c_s + 127) >> 7;
 }
 
-static void calibrate_cpu_clock(void)
+#define NR_TIME_ITERS  50
+
+static int calibrate_cpu_clock(void)
 {
        double delta, mean, S;
-       unsigned long avg, cycles[10];
+       uint64_t 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) {
@@ -230,10 +279,17 @@ static void calibrate_cpu_clock(void)
                }
        }
 
-       S = sqrt(S / (10 - 1.0));
+       /*
+        * 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;
-       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)
@@ -242,91 +298,329 @@ 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++)
-               dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
+       for (i = 0; i < NR_TIME_ITERS; i++)
+               dprint(FD_TIME, "cycles[%d]=%llu\n", i,
+                                       (unsigned long long) cycles[i] / 10);
 
-       avg /= (samples * 10);
-       dprint(FD_TIME, "avg: %lu\n", avg);
+       avg /= samples;
+       avg = (avg + 5) / 10;
+       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 int calibrate_cpu_clock(void)
+{
+       return 1;
+}
+#endif
+
+#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))
+               log_err("fio: can't set TLS key\n");
+}
 
+static void kill_tv_tls_key(void *data)
+{
+       free(data);
 }
 #else
-static void calibrate_cpu_clock(void)
+void fio_local_clock_init(int is_thread)
 {
 }
 #endif
 
 void fio_clock_init(void)
 {
-       last_tv_valid = 0;
-       calibrate_cpu_clock();
+       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_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(struct timeval *s, 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(struct timeval *s)
+{
+       struct timeval t;
+
+       fio_gettime(&t, NULL);
+       return utime_since(s, &t);
 }
 
-void fio_gtod_init(void)
+uint64_t mtime_since(struct timeval *s, struct timeval *e)
 {
-       fio_tv = smalloc(sizeof(struct timeval));
-       assert(fio_tv);
+       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;
 }
 
-void fio_gtod_update(void)
+uint64_t mtime_since_now(struct timeval *s)
 {
-       gettimeofday(fio_tv, NULL);
+       struct timeval t;
+       void *p = __builtin_return_address(0);
+
+       fio_gettime(&t, p);
+       return mtime_since(s, &t);
 }
 
-static void *gtod_thread_main(void *data)
+uint64_t time_since_now(struct timeval *s)
 {
-       struct fio_mutex *mutex = data;
+       return mtime_since_now(s) / 1000;
+}
+
+#if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
+    defined(CONFIG_SFAA)
 
-       fio_mutex_up(mutex);
+#define CLOCK_ENTRIES  100000
+
+struct clock_entry {
+       uint32_t seq;
+       uint32_t cpu;
+       uint64_t tsc;
+};
+
+struct clock_thread {
+       pthread_t thread;
+       int cpu;
+       pthread_mutex_t lock;
+       pthread_mutex_t started;
+       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;
+
+       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);
+
+       last_seq = 0;
+       c = &t->entries[0];
+       for (i = 0; i < CLOCK_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;
+       }
+
+       log_info("cs: cpu%3d: %llu clocks seen\n", t->cpu,
+               (unsigned long long) t->entries[i - 1].tsc - t->entries[0].tsc);
 
        /*
-        * 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.
+        * The most common platform clock breakage is returning zero
+        * indefinitely. Check for that and return failure.
         */
-       while (threads) {
-               fio_gtod_update();
-               nop;
-       }
+       if (!t->entries[i - 1].tsc && !t->entries[0].tsc)
+               return (void *) 1;
 
        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;
+       const struct clock_entry *c1 = p1;
+       const struct clock_entry *c2 = p2;
 
-       mutex = fio_mutex_init(0);
-       if (!mutex)
-               return 1;
+       if (c1->seq == c2->seq)
+               log_err("cs: bug in atomic sequence!\n");
 
-       pthread_attr_init(&attr);
-       pthread_attr_setstacksize(&attr, PTHREAD_STACK_MIN);
-       ret = pthread_create(&gtod_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;
+       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;
+       struct clock_entry *prev, *this;
+       uint32_t seq = 0;
+       int i;
+
+       log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
+
+       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: %lu 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++) {
+               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=%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++;
+               }
+
+               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