Use a monotonic clock instead of a realtime clock

Using a realtime clock is a bad idea: it is affected by any kind of time
change, which can happen when the administrator modifies the system time,
or more simply when a laptop suspends to RAM and then wakes up from sleep.

With the current approach of using a realtime clock:

- if the system time jumps forward (e.g. when resuming after a
  suspend-to-RAM), bmon would take 100% CPU and display random graph data
  extremely fast, until it "catches up" with the new time.

- if the system time jumps backwards, bmon would freeze until *time*
  "catches up" to the point it was before.  bmon then (incorrectly)
  displays a spike in the graph, because lots of packets have been
  sent/received since the last update.

Instead of using gettimeofday(), switch to clock_gettime() with
CLOCK_MONOTONIC on systems that support it.  OS X does not provide
clock_gettime(), so this commit also adds a Mach-specific implementation.

This change has been tested on Linux 4.1 with glibc and musl, and on
FreeBSD 10.0-RELEASE-p12.

src/utils.c

@@ -27,6 +27,11 @@
 #include <bmon/conf.h>
 #include <bmon/utils.h>
 
+#ifdef __MACH__
+#include <mach/clock.h>
+#include <mach/mach.h>
+#endif
+
 void *xcalloc(size_t n, size_t s)
 {
 	void *d = calloc(n, s);
@@ -112,12 +117,21 @@ int timestamp_is_negative(timestamp_t *ts)
 
 void update_timestamp(timestamp_t *dst)
 {
-	struct timeval tv;
+#ifdef __MACH__
+	clock_serv_t cclock;
+	mach_timespec_t tp;
 
-	gettimeofday(&tv, NULL);
+	host_get_clock_service(mach_host_self(), SYSTEM_CLOCK, &cclock);
+	clock_get_time(cclock, &tp);
+	mach_port_deallocate(mach_task_self(), cclock);
+#else
+	struct timespec tp;
 
-	dst->tv_sec = tv.tv_sec;
-	dst->tv_usec = tv.tv_usec;
+	clock_gettime(CLOCK_MONOTONIC, &tp);
+#endif
+
+	dst->tv_sec = tp.tv_sec;
+	dst->tv_usec = tp.tv_nsec / 1000;
 }
 
 void copy_timestamp(timestamp_t *ts1, timestamp_t *ts2)