gettime: offset CPU cycle counter by initial value
authorJens Axboe <axboe@fb.com>
Wed, 17 Dec 2014 15:51:23 +0000 (08:51 -0700)
committerJens Axboe <axboe@fb.com>
Wed, 17 Dec 2014 15:51:23 +0000 (08:51 -0700)
Should then be safe for the full 2^64 cycles, and we push
the more expensive variable division to later in the run (which
is probably never).

Signed-off-by: Jens Axboe <axboe@fb.com>
gettime.c

index 7e3e2e5..a95b545 100644 (file)
--- a/gettime.c
+++ b/gettime.c
@@ -17,6 +17,7 @@
 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;
 
@@ -169,12 +170,16 @@ static void __fio_gettime(struct timeval *tp)
                uint64_t usecs, t;
 
                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");
-                       tv->warned = 1;
+               if (t < cycles_start && !cycles_wrap)
+                       cycles_wrap = 1;
+               else if (cycles_wrap && t >= cycles_start) {
+                       if (!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
@@ -299,6 +304,8 @@ static int calibrate_cpu_clock(void)
        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