gettime: calibration rounding error
[fio.git] / gettime.c
index 48aeb075b65e1c9182e41a3c1240603ebf968691..a60b35f65f2b45df5c129a326ffc2c909afa2bdf 100644 (file)
--- a/gettime.c
+++ b/gettime.c
@@ -1,26 +1,38 @@
 /*
- * Memory helpers
+ * Clock functions
  */
+
 #include <unistd.h>
+#include <math.h>
 #include <sys/time.h>
+#include <time.h>
 
 #include "fio.h"
-#include "os.h"
+#include "smalloc.h"
 
 #include "hash.h"
 
-static int clock_gettime_works = 1;
+#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;
+
+enum fio_cs fio_clock_source = FIO_PREFERRED_CLOCK_SOURCE;
+int fio_clock_source_set = 0;
 
 #ifdef FIO_DEBUG_TIME
 
 #define HASH_BITS      8
 #define HASH_SIZE      (1 << HASH_BITS)
 
-static struct list_head hash[HASH_SIZE];
+static struct flist_head hash[HASH_SIZE];
 static int gtod_inited;
 
 struct gtod_log {
-       struct list_head list;
+       struct flist_head list;
        void *caller;
        unsigned long calls;
 };
@@ -28,10 +40,11 @@ struct gtod_log {
 static struct gtod_log *find_hash(void *caller)
 {
        unsigned long h = hash_ptr(caller, HASH_BITS);
-       struct list_head *entry;
+       struct flist_head *entry;
 
-       list_for_each(entry, &hash[h]) {
-               struct gtod_log *log = list_entry(entry, struct gtod_log, list);
+       flist_for_each(entry, &hash[h]) {
+               struct gtod_log *log = flist_entry(entry, struct gtod_log,
+                                                                       list);
 
                if (log->caller == caller)
                        return log;
@@ -48,12 +61,12 @@ static struct gtod_log *find_log(void *caller)
                unsigned long h;
 
                log = malloc(sizeof(*log));
-               INIT_LIST_HEAD(&log->list);
+               INIT_FLIST_HEAD(&log->list);
                log->caller = caller;
                log->calls = 0;
 
                h = hash_ptr(caller, HASH_BITS);
-               list_add_tail(&log->list, &hash[h]);
+               flist_add_tail(&log->list, &hash[h]);
        }
 
        return log;
@@ -74,13 +87,14 @@ static void fio_exit fio_dump_gtod(void)
        int i;
 
        for (i = 0; i < HASH_SIZE; i++) {
-               struct list_head *entry;
+               struct flist_head *entry;
                struct gtod_log *log;
 
-               list_for_each(entry, &hash[i]) {
-                       log = list_entry(entry, struct gtod_log, list);
+               flist_for_each(entry, &hash[i]) {
+                       log = flist_entry(entry, struct gtod_log, list);
 
-                       printf("function %p, calls %lu\n", log->caller, log->calls);
+                       printf("function %p, calls %lu\n", log->caller,
+                                                               log->calls);
                        total_calls += log->calls;
                }
        }
@@ -93,43 +107,237 @@ static void fio_init gtod_init(void)
        int i;
 
        for (i = 0; i < HASH_SIZE; i++)
-               INIT_LIST_HEAD(&hash[i]);
+               INIT_FLIST_HEAD(&hash[i]);
 
        gtod_inited = 1;
 }
 
 #endif /* FIO_DEBUG_TIME */
 
+#ifdef FIO_DEBUG_TIME
 void fio_gettime(struct timeval *tp, void *caller)
+#else
+void fio_gettime(struct timeval *tp, void fio_unused *caller)
+#endif
 {
 #ifdef FIO_DEBUG_TIME
        if (!caller)
                caller = __builtin_return_address(0);
 
        gtod_log_caller(caller);
-#else
-       caller = NULL;
 #endif
-repeat:
-       if (!clock_gettime_works)
+       if (fio_tv) {
+               memcpy(tp, fio_tv, sizeof(*tp));
+               return;
+       }
+
+       switch (fio_clock_source) {
+       case CS_GTOD:
                gettimeofday(tp, NULL);
-       else {
+               break;
+       case CS_CGETTIME: {
                struct timespec ts;
 
+#ifdef FIO_HAVE_CLOCK_MONOTONIC
                if (clock_gettime(CLOCK_MONOTONIC, &ts) < 0) {
-                       clock_gettime_works = 0;
-                       goto repeat;
+#else
+               if (clock_gettime(CLOCK_REALTIME, &ts) < 0) {
+#endif
+                       log_err("fio: clock_gettime fails\n");
+                       assert(0);
                }
 
                tp->tv_sec = ts.tv_sec;
                tp->tv_usec = ts.tv_nsec / 1000;
+               break;
+               }
+#ifdef ARCH_HAVE_CPU_CLOCK
+       case CS_CPUCLOCK: {
+               unsigned long long usecs, t;
+
+               t = get_cpu_clock();
+               if (t < last_cycles) {
+                       dprint(FD_TIME, "CPU clock going back in time\n");
+                       t = last_cycles;
+               }
+
+               usecs = t / cycles_per_usec;
+               tp->tv_sec = usecs / 1000000;
+               tp->tv_usec = usecs % 1000000;
+               last_cycles = t;
+               break;
+               }
+#endif
+       default:
+               log_err("fio: invalid clock source %d\n", fio_clock_source);
+               break;
+       }
+
+       /*
+        * 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;
+       }
+       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;
+
+       gettimeofday(&s, NULL);
+       c_s = get_cpu_clock();
+       do {
+               unsigned long long elapsed;
+
+               gettimeofday(&e, NULL);
+               elapsed = utime_since(&s, &e);
+               if (elapsed >= 1280) {
+                       c_e = get_cpu_clock();
+                       break;
+               }
+       } while (1);
+
+       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[NR_TIME_ITERS];
+       int i, samples;
+
+       cycles[0] = get_cycles_per_usec();
+       S = delta = mean = 0.0;
+       for (i = 0; i < NR_TIME_ITERS; i++) {
+               cycles[i] = get_cycles_per_usec();
+               delta = cycles[i] - mean;
+               if (delta) {
+                       mean += delta / (i + 1.0);
+                       S += delta * (cycles[i] - mean);
+               }
+       }
+
+       S = sqrt(S / (NR_TIME_ITERS - 1.0));
+
+       samples = avg = 0;
+       for (i = 0; i < NR_TIME_ITERS; i++) {
+               double this = cycles[i];
+
+               if ((fmax(this, mean) - fmin(this, mean)) > S)
+                       continue;
+               samples++;
+               avg += this;
        }
+
+       S /= (double) NR_TIME_ITERS;
+       mean /= (double) NR_TIME_ITERS;
+
+       for (i = 0; i < NR_TIME_ITERS; i++)
+               dprint(FD_TIME, "cycles[%d]=%lu\n", i, cycles[i] / 10);
+
+       avg /= (samples * 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)
+{
+}
+#endif
 
-static void fio_init check_clock(void)
+void fio_clock_init(void)
 {
-       struct timespec ts;
+       last_tv_valid = 0;
+       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");
+}
 
-       if (clock_getres(CLOCK_MONOTONIC, &ts) < 0)
-               clock_gettime_works = 0;
+unsigned long long utime_since(struct timeval *s, struct timeval *e)
+{
+       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;
+}
+
+unsigned long long utime_since_now(struct timeval *s)
+{
+       struct timeval t;
+
+       fio_gettime(&t, NULL);
+       return utime_since(s, &t);
+}
+
+unsigned long mtime_since(struct timeval *s, 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;
+}
+
+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;
 }