gettime: add basic init cpuclock test
[fio.git] / gettime.c
index 87abca0ff883f59c4850f191ae8d744128552bee..5ceb809de3ae9247a5cc7d3e2557b062541fc73f 100644 (file)
--- a/gettime.c
+++ b/gettime.c
@@ -16,6 +16,8 @@
 #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;
+static uint64_t max_cycles_for_mult;
+static unsigned long long cycles_start, cycles_wrap;
 #endif
 int tsc_reliable = 0;
 
@@ -24,11 +26,13 @@ struct tv_valid {
        int last_tv_valid;
        int warned;
 };
+#ifdef ARCH_HAVE_CPU_CLOCK
 #ifdef CONFIG_TLS_THREAD
 static __thread struct tv_valid static_tv_valid;
 #else
 static pthread_key_t tv_tls_key;
 #endif
+#endif
 
 enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
 int fio_clock_source_set = 0;
@@ -64,7 +68,7 @@ static struct gtod_log *find_hash(void *caller)
        return NULL;
 }
 
-static struct gtod_log *find_log(void *caller)
+static void inc_caller(void *caller)
 {
        struct gtod_log *log = find_hash(caller);
 
@@ -80,16 +84,13 @@ static struct gtod_log *find_log(void *caller)
                flist_add_tail(&log->list, &hash[h]);
        }
 
-       return log;
+       log->calls++;
 }
 
 static void gtod_log_caller(void *caller)
 {
-       if (gtod_inited) {
-               struct gtod_log *log = find_log(caller);
-
-               log->calls++;
-       }
+       if (gtod_inited)
+               inc_caller(caller);
 }
 
 static void fio_exit fio_dump_gtod(void)
@@ -138,14 +139,6 @@ static int fill_clock_gettime(struct timespec *ts)
 
 static void __fio_gettime(struct timeval *tp)
 {
-       struct tv_valid *tv;
-
-#ifdef CONFIG_TLS_THREAD
-       tv = &static_tv_valid;
-#else
-       tv = pthread_getspecific(tv_tls_key);
-#endif
-
        switch (fio_clock_source) {
 #ifdef CONFIG_GETTIMEOFDAY
        case CS_GTOD:
@@ -169,20 +162,32 @@ static void __fio_gettime(struct timeval *tp)
 #ifdef ARCH_HAVE_CPU_CLOCK
        case CS_CPUCLOCK: {
                uint64_t usecs, t;
+               struct tv_valid *tv;
+
+#ifdef CONFIG_TLS_THREAD
+               tv = &static_tv_valid;
+#else
+               tv = pthread_getspecific(tv_tls_key);
+#endif
 
                t = get_cpu_clock();
-               if (t < tv->last_cycles && tv->last_tv_valid &&
-                   !tv->warned) {
-                       log_err("fio: CPU clock going back in time\n");
+               if (t < cycles_start && !cycles_wrap)
+                       cycles_wrap = 1;
+               else if (cycles_wrap && t >= cycles_start && !tv->warned) {
+                       log_err("fio: double CPU clock wrap\n");
                        tv->warned = 1;
                }
 
+               t -= cycles_start;
                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;
+               if (t < max_cycles_for_mult)
+                       usecs = (t * inv_cycles_per_usec) / 16777216UL;
+               else
+                       usecs = t / cycles_per_usec;
 #endif
                tp->tv_sec = usecs / 1000000;
                tp->tv_usec = usecs % 1000000;
@@ -207,10 +212,8 @@ void fio_gettime(struct timeval *tp, void fio_unused *caller)
 
        gtod_log_caller(caller);
 #endif
-       if (fio_unlikely(fio_tv)) {
-               memcpy(tp, fio_tv, sizeof(*tp));
+       if (fio_unlikely(fio_gettime_offload(tp)))
                return;
-       }
 
        __fio_gettime(tp);
 }
@@ -298,7 +301,10 @@ static int calibrate_cpu_clock(void)
 
        cycles_per_usec = avg;
        inv_cycles_per_usec = 16777216UL / cycles_per_usec;
+       max_cycles_for_mult = ~0ULL / inv_cycles_per_usec;
        dprint(FD_TIME, "inv_cycles_per_usec=%lu\n", inv_cycles_per_usec);
+       cycles_start = get_cpu_clock();
+       dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
        return 0;
 }
 #else
@@ -355,7 +361,7 @@ void fio_clock_init(void)
         * runs at a constant rate and is synced across CPU cores.
         */
        if (tsc_reliable) {
-               if (!fio_clock_source_set)
+               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");
@@ -430,7 +436,8 @@ uint64_t time_since_now(const struct timeval *s)
 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
     defined(CONFIG_SFAA)
 
-#define CLOCK_ENTRIES  100000
+#define CLOCK_ENTRIES_DEBUG    100000
+#define CLOCK_ENTRIES_TEST     10000
 
 struct clock_entry {
        uint32_t seq;
@@ -441,8 +448,10 @@ struct clock_entry {
 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;
 };
@@ -473,7 +482,7 @@ static void *clock_thread_fn(void *data)
 
        last_seq = 0;
        c = &t->entries[0];
-       for (i = 0; i < CLOCK_ENTRIES; i++, c++) {
+       for (i = 0; i < t->nr_entries; i++, c++) {
                uint32_t seq;
                uint64_t tsc;
 
@@ -489,8 +498,12 @@ static void *clock_thread_fn(void *data)
                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);
+       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
@@ -513,38 +526,49 @@ static int clock_cmp(const void *p1, const void *p2)
        return c1->seq - c2->seq;
 }
 
-int fio_monotonic_clocktest(void)
+int fio_monotonic_clocktest(int debug)
 {
        struct clock_thread *cthreads;
        unsigned int nr_cpus = cpus_online();
        struct clock_entry *entries;
-       unsigned long tentries, failed = 0;
+       unsigned long nr_entries, 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");
+       if (debug) {
+               log_info("cs: reliable_tsc: %s\n", tsc_reliable ? "yes" : "no");
 
 #ifdef FIO_INC_DEBUG
-       fio_debug |= 1U << FD_TIME;
+               fio_debug |= 1U << FD_TIME;
 #endif
+               nr_entries = CLOCK_ENTRIES_DEBUG;
+       } else
+               nr_entries = CLOCK_ENTRIES_TEST;
+
        calibrate_cpu_clock();
+
+       if (debug) {
 #ifdef FIO_INC_DEBUG
-       fio_debug &= ~(1U << FD_TIME);
+               fio_debug &= ~(1U << FD_TIME);
 #endif
+       }
 
        cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
-       tentries = CLOCK_ENTRIES * nr_cpus;
+       tentries = nr_entries * nr_cpus;
        entries = malloc(tentries * sizeof(struct clock_entry));
 
-       log_info("cs: Testing %u CPUs\n", nr_cpus);
+       if (debug)
+               log_info("cs: Testing %u CPUs\n", nr_cpus);
 
        for (i = 0; i < nr_cpus; i++) {
                struct clock_thread *t = &cthreads[i];
 
                t->cpu = i;
+               t->debug = debug;
                t->seq = &seq;
-               t->entries = &entries[i * CLOCK_ENTRIES];
+               t->nr_entries = nr_entries;
+               t->entries = &entries[i * nr_entries];
                pthread_mutex_init(&t->lock, NULL);
                pthread_mutex_init(&t->started, NULL);
                pthread_mutex_lock(&t->lock);
@@ -578,7 +602,8 @@ int fio_monotonic_clocktest(void)
        free(cthreads);
 
        if (failed) {
-               log_err("Clocksource test: %lu threads failed\n", failed);
+               if (debug)
+                       log_err("Clocksource test: %lu threads failed\n", failed);
                goto err;
        }
 
@@ -595,6 +620,11 @@ int fio_monotonic_clocktest(void)
                if (prev->tsc > this->tsc) {
                        uint64_t diff = prev->tsc - this->tsc;
 
+                       if (!debug) {
+                               failed++;
+                               break;
+                       }
+
                        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);
@@ -605,11 +635,12 @@ int fio_monotonic_clocktest(void)
                prev = this;
        }
 
-       if (failed)
-               log_info("cs: Failed: %lu\n", failed);
-       else
-               log_info("cs: Pass!\n");
-
+       if (debug) {
+               if (failed)
+                       log_info("cs: Failed: %lu\n", failed);
+               else
+                       log_info("cs: Pass!\n");
+       }
 err:
        free(entries);
        return !!failed;
@@ -617,10 +648,11 @@ err:
 
 #else /* defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK) */
 
-int fio_monotonic_clocktest(void)
+int fio_monotonic_clocktest(int debug)
 {
-       log_info("cs: current platform does not support CPU clocks\n");
-       return 0;
+       if (debug)
+               log_info("cs: current platform does not support CPU clocks\n");
+       return 1;
 }
 
 #endif