configure: attempt to link against tcmalloc by default if available
[fio.git] / gettime.c
index d261e22e968b864ab1723b819759e87024d1cc9f..272a3e6244527c27d59a06ff99fb4871424a063b 100644 (file)
--- a/gettime.c
+++ b/gettime.c
@@ -2,33 +2,38 @@
  * Clock functions
  */
 
-#include <unistd.h>
 #include <math.h>
-#include <sys/time.h>
-#include <time.h>
 
 #include "fio.h"
-#include "smalloc.h"
-
-#include "hash.h"
 #include "os/os.h"
 
-#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;
+#if defined(ARCH_HAVE_CPU_CLOCK)
+#ifndef ARCH_CPU_CLOCK_CYCLES_PER_USEC
+static unsigned long long cycles_per_msec;
+static unsigned long long cycles_start;
+static unsigned long long clock_mult;
+static unsigned long long max_cycles_mask;
+static unsigned long long nsecs_for_max_cycles;
+static unsigned int clock_shift;
+static unsigned int max_cycles_shift;
+#define MAX_CLOCK_SEC 60*60
+#endif
+#ifdef ARCH_CPU_CLOCK_WRAPS
+static unsigned int cycles_wrap;
+#endif
 #endif
-int tsc_reliable = 0;
+bool tsc_reliable = false;
 
 struct tv_valid {
-       struct timeval last_tv;
-       uint64_t last_cycles;
-       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 +69,7 @@ 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 +85,13 @@ 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)
@@ -128,7 +130,9 @@ static void fio_init gtod_init(void)
 #ifdef CONFIG_CLOCK_GETTIME
 static int fill_clock_gettime(struct timespec *ts)
 {
-#ifdef CONFIG_CLOCK_MONOTONIC
+#if defined(CONFIG_CLOCK_MONOTONIC_RAW)
+       return clock_gettime(CLOCK_MONOTONIC_RAW, ts);
+#elif defined(CONFIG_CLOCK_MONOTONIC)
        return clock_gettime(CLOCK_MONOTONIC, ts);
 #else
        return clock_gettime(CLOCK_REALTIME, ts);
@@ -136,54 +140,58 @@ static int fill_clock_gettime(struct timespec *ts)
 }
 #endif
 
-static void *__fio_gettime(struct timeval *tp)
+static void __fio_gettime(struct timespec *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:
-               gettimeofday(tp, NULL);
+       case CS_GTOD: {
+               struct timeval tv;
+               gettimeofday(&tv, NULL);
+
+               tp->tv_sec = tv.tv_sec;
+               tp->tv_nsec = tv.tv_usec * 1000;
                break;
+               }
 #endif
 #ifdef CONFIG_CLOCK_GETTIME
        case CS_CGETTIME: {
-               struct timespec ts;
-
-               if (fill_clock_gettime(&ts) < 0) {
+               if (fill_clock_gettime(tp) < 0) {
                        log_err("fio: clock_gettime fails\n");
                        assert(0);
                }
-
-               tp->tv_sec = ts.tv_sec;
-               tp->tv_usec = ts.tv_nsec / 1000;
                break;
                }
 #endif
 #ifdef ARCH_HAVE_CPU_CLOCK
        case CS_CPUCLOCK: {
-               uint64_t usecs, t;
+               uint64_t nsecs, t, multiples;
+               struct tv_valid *tv;
 
-               t = get_cpu_clock();
-               if (tv && t < tv->last_cycles) {
-                       dprint(FD_TIME, "CPU clock going back in time\n");
-                       t = tv->last_cycles;
-               } else if (tv)
-                       tv->last_cycles = t;
+#ifdef CONFIG_TLS_THREAD
+               tv = &static_tv_valid;
+#else
+               tv = pthread_getspecific(tv_tls_key);
+#endif
 
+               t = get_cpu_clock();
+#ifdef ARCH_CPU_CLOCK_WRAPS
+               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;
+               }
+#endif
 #ifdef ARCH_CPU_CLOCK_CYCLES_PER_USEC
-               usecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC;
+               nsecs = t / ARCH_CPU_CLOCK_CYCLES_PER_USEC * 1000;
 #else
-               usecs = (t * inv_cycles_per_usec) / 16777216UL;
+               t -= cycles_start;
+               multiples = t >> max_cycles_shift;
+               nsecs = multiples * nsecs_for_max_cycles;
+               nsecs += ((t & max_cycles_mask) * clock_mult) >> clock_shift;
 #endif
-               tp->tv_sec = usecs / 1000000;
-               tp->tv_usec = usecs % 1000000;
+               tp->tv_sec = nsecs / 1000000000ULL;
+               tp->tv_nsec = nsecs % 1000000000ULL;
                break;
                }
 #endif
@@ -191,54 +199,33 @@ static void *__fio_gettime(struct timeval *tp)
                log_err("fio: invalid clock source %d\n", fio_clock_source);
                break;
        }
-
-       return tv;
 }
 
 #ifdef FIO_DEBUG_TIME
-void fio_gettime(struct timeval *tp, void *caller)
+void fio_gettime(struct timespec *tp, void *caller)
 #else
-void fio_gettime(struct timeval *tp, void fio_unused *caller)
+void fio_gettime(struct timespec *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_unlikely(fio_tv)) {
-               memcpy(tp, fio_tv, sizeof(*tp));
+       if (fio_unlikely(fio_gettime_offload(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 (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));
-       }
+       __fio_gettime(tp);
 }
 
 #if defined(ARCH_HAVE_CPU_CLOCK) && !defined(ARCH_CPU_CLOCK_CYCLES_PER_USEC)
-static unsigned long get_cycles_per_usec(void)
+static unsigned long get_cycles_per_msec(void)
 {
-       struct timeval s, e;
+       struct timespec s, e;
        uint64_t c_s, c_e;
        enum fio_cs old_cs = fio_clock_source;
+       uint64_t elapsed;
 
 #ifdef CONFIG_CLOCK_GETTIME
        fio_clock_source = CS_CGETTIME;
@@ -249,19 +236,16 @@ static unsigned long get_cycles_per_usec(void)
 
        c_s = get_cpu_clock();
        do {
-               uint64_t elapsed;
-
                __fio_gettime(&e);
+               c_e = get_cpu_clock();
 
-               elapsed = utime_since(&s, &e);
-               if (elapsed >= 1280) {
-                       c_e = get_cpu_clock();
+               elapsed = ntime_since(&s, &e);
+               if (elapsed >= 1280000)
                        break;
-               }
        } while (1);
 
        fio_clock_source = old_cs;
-       return (c_e - c_s + 127) >> 7;
+       return (c_e - c_s) * 1000000 / elapsed;
 }
 
 #define NR_TIME_ITERS  50
@@ -269,13 +253,14 @@ static unsigned long get_cycles_per_usec(void)
 static int calibrate_cpu_clock(void)
 {
        double delta, mean, S;
-       uint64_t avg, cycles[NR_TIME_ITERS];
-       int i, samples;
+       uint64_t minc, maxc, avg, cycles[NR_TIME_ITERS];
+       int i, samples, sft = 0;
+       unsigned long long tmp, max_ticks, max_mult;
 
-       cycles[0] = get_cycles_per_usec();
+       cycles[0] = get_cycles_per_msec();
        S = delta = mean = 0.0;
        for (i = 0; i < NR_TIME_ITERS; i++) {
-               cycles[i] = get_cycles_per_usec();
+               cycles[i] = get_cycles_per_msec();
                delta = cycles[i] - mean;
                if (delta) {
                        mean += delta / (i + 1.0);
@@ -292,10 +277,14 @@ static int calibrate_cpu_clock(void)
 
        S = sqrt(S / (NR_TIME_ITERS - 1.0));
 
-       samples = avg = 0;
+       minc = -1ULL;
+       maxc = samples = avg = 0;
        for (i = 0; i < NR_TIME_ITERS; i++) {
                double this = cycles[i];
 
+               minc = min(cycles[i], minc);
+               maxc = max(cycles[i], maxc);
+
                if ((fmax(this, mean) - fmin(this, mean)) > S)
                        continue;
                samples++;
@@ -303,20 +292,72 @@ static int calibrate_cpu_clock(void)
        }
 
        S /= (double) NR_TIME_ITERS;
-       mean /= 10.0;
 
        for (i = 0; i < NR_TIME_ITERS; i++)
-               dprint(FD_TIME, "cycles[%d]=%llu\n", i,
-                                       (unsigned long long) cycles[i] / 10);
+               dprint(FD_TIME, "cycles[%d]=%llu\n", i, (unsigned long long) cycles[i]);
 
        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_msec = avg;
+       dprint(FD_TIME, "min=%llu, max=%llu, mean=%f, S=%f, N=%d\n",
+                       (unsigned long long) minc,
+                       (unsigned long long) maxc, mean, S, NR_TIME_ITERS);
+       dprint(FD_TIME, "trimmed mean=%llu, N=%d\n", (unsigned long long) avg, samples);
+
+       max_ticks = MAX_CLOCK_SEC * cycles_per_msec * 1000ULL;
+       max_mult = ULLONG_MAX / max_ticks;
+       dprint(FD_TIME, "\n\nmax_ticks=%llu, __builtin_clzll=%d, "
+                       "max_mult=%llu\n", max_ticks,
+                       __builtin_clzll(max_ticks), max_mult);
+
+        /*
+         * Find the largest shift count that will produce
+         * a multiplier that does not exceed max_mult
+         */
+        tmp = max_mult * cycles_per_msec / 1000000;
+        while (tmp > 1) {
+                tmp >>= 1;
+                sft++;
+                dprint(FD_TIME, "tmp=%llu, sft=%u\n", tmp, sft);
+        }
+
+       clock_shift = sft;
+       clock_mult = (1ULL << sft) * 1000000 / cycles_per_msec;
+       dprint(FD_TIME, "clock_shift=%u, clock_mult=%llu\n", clock_shift,
+                                                       clock_mult);
 
-       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);
+       /*
+        * Find the greatest power of 2 clock ticks that is less than the
+        * ticks in MAX_CLOCK_SEC_2STAGE
+        */
+       max_cycles_shift = max_cycles_mask = 0;
+       tmp = MAX_CLOCK_SEC * 1000ULL * cycles_per_msec;
+       dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp,
+                                                       max_cycles_shift);
+       while (tmp > 1) {
+               tmp >>= 1;
+               max_cycles_shift++;
+               dprint(FD_TIME, "tmp=%llu, max_cycles_shift=%u\n", tmp, max_cycles_shift);
+       }
+       /*
+        * if use use (1ULL << max_cycles_shift) * 1000 / cycles_per_msec
+        * here we will have a discontinuity every
+        * (1ULL << max_cycles_shift) cycles
+        */
+       nsecs_for_max_cycles = ((1ULL << max_cycles_shift) * clock_mult)
+                                       >> clock_shift;
+
+       /* Use a bitmask to calculate ticks % (1ULL << max_cycles_shift) */
+       for (tmp = 0; tmp < max_cycles_shift; tmp++)
+               max_cycles_mask |= 1ULL << tmp;
+
+       dprint(FD_TIME, "max_cycles_shift=%u, 2^max_cycles_shift=%llu, "
+                       "nsecs_for_max_cycles=%llu, "
+                       "max_cycles_mask=%016llx\n",
+                       max_cycles_shift, (1ULL << max_cycles_shift),
+                       nsecs_for_max_cycles, max_cycles_mask);
+
+       cycles_start = get_cpu_clock();
+       dprint(FD_TIME, "cycles_start=%llu\n", cycles_start);
        return 0;
 }
 #else
@@ -331,13 +372,15 @@ static int calibrate_cpu_clock(void)
 #endif // ARCH_HAVE_CPU_CLOCK
 
 #ifndef CONFIG_TLS_THREAD
-void fio_local_clock_init(int is_thread)
+void fio_local_clock_init(void)
 {
        struct tv_valid *t;
 
        t = calloc(1, sizeof(*t));
-       if (pthread_setspecific(tv_tls_key, t))
+       if (pthread_setspecific(tv_tls_key, t)) {
                log_err("fio: can't set TLS key\n");
+               assert(0);
+       }
 }
 
 static void kill_tv_tls_key(void *data)
@@ -345,7 +388,7 @@ static void kill_tv_tls_key(void *data)
        free(data);
 }
 #else
-void fio_local_clock_init(int is_thread)
+void fio_local_clock_init(void)
 {
 }
 #endif
@@ -363,7 +406,7 @@ void fio_clock_init(void)
        fio_clock_source_inited = fio_clock_source;
 
        if (calibrate_cpu_clock())
-               tsc_reliable = 0;
+               tsc_reliable = false;
 
        /*
         * If the arch sets tsc_reliable != 0, then it must be good enough
@@ -371,19 +414,47 @@ 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");
+       dprint(FD_TIME, "gettime: clocksource=%d\n", (int) fio_clock_source);
+}
+
+uint64_t ntime_since(const struct timespec *s, const struct timespec *e)
+{
+       int64_t sec, nsec;
+
+       sec = e->tv_sec - s->tv_sec;
+       nsec = e->tv_nsec - s->tv_nsec;
+       if (sec > 0 && nsec < 0) {
+              sec--;
+              nsec += 1000000000LL;
+       }
+
+       /*
+       * time warp bug on some kernels?
+       */
+       if (sec < 0 || (sec == 0 && nsec < 0))
+              return 0;
+
+       return nsec + (sec * 1000000000LL);
+}
+
+uint64_t ntime_since_now(const struct timespec *s)
+{
+       struct timespec now;
+
+       fio_gettime(&now, NULL);
+       return ntime_since(s, &now);
 }
 
-uint64_t utime_since(struct timeval *s, struct timeval *e)
+uint64_t utime_since(const struct timespec *s, const struct timespec *e)
 {
-       long sec, usec;
-       uint64_t ret;
+       int64_t sec, usec;
 
        sec = e->tv_sec - s->tv_sec;
-       usec = e->tv_usec - s->tv_usec;
+       usec = (e->tv_nsec - s->tv_nsec) / 1000;
        if (sec > 0 && usec < 0) {
                sec--;
                usec += 1000000;
@@ -395,25 +466,29 @@ uint64_t utime_since(struct timeval *s, struct timeval *e)
        if (sec < 0 || (sec == 0 && usec < 0))
                return 0;
 
-       ret = sec * 1000000ULL + usec;
-
-       return ret;
+       return usec + (sec * 1000000);
 }
 
-uint64_t utime_since_now(struct timeval *s)
+uint64_t utime_since_now(const struct timespec *s)
 {
-       struct timeval t;
+       struct timespec t;
+#ifdef FIO_DEBUG_TIME
+       void *p = __builtin_return_address(0);
 
+       fio_gettime(&t, p);
+#else
        fio_gettime(&t, NULL);
+#endif
+
        return utime_since(s, &t);
 }
 
-uint64_t mtime_since(struct timeval *s, struct timeval *e)
+uint64_t mtime_since_tv(const struct timeval *s, const struct timeval *e)
 {
-       long sec, usec, ret;
+       int64_t sec, usec;
 
        sec = e->tv_sec - s->tv_sec;
-       usec = e->tv_usec - s->tv_usec;
+       usec = (e->tv_usec - s->tv_usec);
        if (sec > 0 && usec < 0) {
                sec--;
                usec += 1000000;
@@ -422,31 +497,54 @@ uint64_t mtime_since(struct timeval *s, struct timeval *e)
        if (sec < 0 || (sec == 0 && usec < 0))
                return 0;
 
-       sec *= 1000UL;
-       usec /= 1000UL;
-       ret = sec + usec;
-
-       return ret;
+       sec *= 1000;
+       usec /= 1000;
+       return sec + usec;
 }
 
-uint64_t mtime_since_now(struct timeval *s)
+uint64_t mtime_since_now(const struct timespec *s)
 {
-       struct timeval t;
+       struct timespec t;
+#ifdef FIO_DEBUG_TIME
        void *p = __builtin_return_address(0);
 
        fio_gettime(&t, p);
+#else
+       fio_gettime(&t, NULL);
+#endif
+
        return mtime_since(s, &t);
 }
 
-uint64_t time_since_now(struct timeval *s)
+uint64_t mtime_since(const struct timespec *s, const struct timespec *e)
+{
+       int64_t sec, usec;
+
+       sec = e->tv_sec - s->tv_sec;
+       usec = (e->tv_nsec - s->tv_nsec) / 1000;
+       if (sec > 0 && usec < 0) {
+               sec--;
+               usec += 1000000;
+       }
+
+       if (sec < 0 || (sec == 0 && usec < 0))
+               return 0;
+
+       sec *= 1000;
+       usec /= 1000;
+       return sec + usec;
+}
+
+uint64_t time_since_now(const struct timespec *s)
 {
        return mtime_since_now(s) / 1000;
 }
 
 #if defined(FIO_HAVE_CPU_AFFINITY) && defined(ARCH_HAVE_CPU_CLOCK)  && \
-    defined(CONFIG_SFAA)
+    defined(CONFIG_SYNC_SYNC) && defined(CONFIG_CMP_SWAP)
 
-#define CLOCK_ENTRIES  100000
+#define CLOCK_ENTRIES_DEBUG    100000
+#define CLOCK_ENTRIES_TEST     1000
 
 struct clock_entry {
        uint32_t seq;
@@ -457,15 +555,17 @@ struct clock_entry {
 struct clock_thread {
        pthread_t thread;
        int cpu;
-       pthread_mutex_t lock;
-       pthread_mutex_t started;
+       int debug;
+       struct fio_sem lock;
+       unsigned long nr_entries;
        uint32_t *seq;
        struct clock_entry *entries;
 };
 
-static inline uint32_t atomic32_inc_return(uint32_t *seq)
+static inline uint32_t atomic32_compare_and_swap(uint32_t *ptr, uint32_t old,
+                                                uint32_t new)
 {
-       return 1 + __sync_fetch_and_add(seq, 1);
+       return __sync_val_compare_and_swap(ptr, old, new);
 }
 
 static void *clock_thread_fn(void *data)
@@ -473,49 +573,70 @@ 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;
+       unsigned long long first;
        int i;
 
-       memset(&cpu_mask, 0, sizeof(cpu_mask));
+       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) {
-               log_err("clock setaffinity failed\n");
-               return (void *) 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);
+       fio_sem_down(&t->lock);
 
-       last_seq = 0;
+       first = get_cpu_clock();
        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;
 
                c->cpu = t->cpu;
                do {
-                       seq = atomic32_inc_return(t->seq);
-                       if (seq < last_seq)
+                       seq = *t->seq;
+                       if (seq == UINT_MAX)
                                break;
+                       __sync_synchronize();
                        tsc = get_cpu_clock();
-               } while (seq != *t->seq);
+               } while (seq != atomic32_compare_and_swap(t->seq, seq, seq + 1));
+
+               if (seq == UINT_MAX)
+                       break;
 
                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);
+       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, first %llu\n", t->cpu,
+                                                       clocks, first);
+       }
 
        /*
         * 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)
-               return (void *) 1;
+       if (i > 1 && !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)
@@ -529,37 +650,50 @@ 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 *threads;
+       struct clock_thread *cthreads;
        unsigned int nr_cpus = cpus_online();
        struct clock_entry *entries;
-       unsigned long tentries, failed;
+       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;
+#endif
+               nr_entries = CLOCK_ENTRIES_DEBUG;
+       } else
+               nr_entries = CLOCK_ENTRIES_TEST;
 
-       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;
+       if (debug) {
+#ifdef FIO_INC_DEBUG
+               fio_debug &= ~(1U << FD_TIME);
+#endif
+       }
+
+       cthreads = malloc(nr_cpus * sizeof(struct clock_thread));
+       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 = &threads[i];
+               struct clock_thread *t = &cthreads[i];
 
                t->cpu = i;
+               t->debug = debug;
                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);
+               t->nr_entries = nr_entries;
+               t->entries = &entries[i * nr_entries];
+               __fio_sem_init(&t->lock, FIO_SEM_LOCKED);
                if (pthread_create(&t->thread, NULL, clock_thread_fn, t)) {
                        failed++;
                        nr_cpus = i;
@@ -568,34 +702,32 @@ int fio_monotonic_clocktest(void)
        }
 
        for (i = 0; i < nr_cpus; i++) {
-               struct clock_thread *t = &threads[i];
+               struct clock_thread *t = &cthreads[i];
 
-               pthread_mutex_lock(&t->started);
+               fio_sem_up(&t->lock);
        }
 
        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];
+               struct clock_thread *t = &cthreads[i];
                void *ret;
 
                pthread_join(t->thread, &ret);
                if (ret)
                        failed++;
+               __fio_sem_remove(&t->lock);
        }
-       free(threads);
+       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;
        }
 
        qsort(entries, tentries, sizeof(struct clock_entry), clock_cmp);
 
+       /* silence silly gcc */
+       prev = NULL;
        for (failed = i = 0; i < tentries; i++) {
                this = &entries[i];
 
@@ -607,6 +739,11 @@ 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);
@@ -617,11 +754,12 @@ 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;
@@ -629,10 +767,11 @@ 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