Description:
------------
get_cpu_frequency() calibrates RDTSC values by measuring the delta across a usleep(5000) call. On my Intel processor on Linux 3.5, the processor often halts for the majority of those 5 milliseconds. The RDTSC counter is not incremented during the halt. So the resulting calibration is off by a factor of 10 or so.

The solution is to do a busy loop instead of a usleep() call. Patch attached. The volatile loop counter prevents the loop from being optimised away, I checked the assembly generated by gcc.

I confirmed that with the patch, the reported CPU frequency is consistent with the physical clock frequency, and stable with a standard deviation of 0.007%.

Test script:
---------------
diff --git a/extension/xhprof.c b/extension/xhprof.c
index a053ede..52841d1 100644
--- a/extension/xhprof.c
+++ b/extension/xhprof.c
@@ -1351,6 +1351,7 @@ static void get_all_cpu_frequencies() {
     usleep(0);
 
     frequency = get_cpu_frequency();
+    printf("CPU %d: %g\n", id, frequency);
     if (frequency == 0.0) {
       clear_frequencies();
       return;