trace: add high resolution timer function to debug performance issues

Add a getnanotime() function that returns nanoseconds since 01/01/1970 as
unsigned 64-bit integer (i.e. overflows in july 2554). This is easier to
work with than e.g. struct timeval or struct timespec. Basing the timer on
the epoch allows using the results with other time-related APIs.

To simplify adaption to different platforms, split the implementation into
a common getnanotime() and a platform-specific highres_nanos() function.

The common getnanotime() function handles errors, falling back to
gettimeofday() if highres_nanos() isn't implemented or doesn't work.

getnanotime() is also responsible for normalizing to the epoch. The offset
to the system clock is calculated only once on initialization, i.e.
manually setting the system clock has no impact on the timer (except if
the fallback gettimeofday() is in use). Git processes are typically short
lived, so we don't need to handle clock drift.

The highres_nanos() function returns monotonically increasing nanoseconds
relative to some arbitrary point in time (e.g. system boot), or 0 on
failure. Providing platform-specific implementations should be relatively
easy, e.g. adapting to clock_gettime() as defined by the POSIX realtime
extensions is seven lines of code.

This version includes highres_nanos() implementations for:
 * Linux: using clock_gettime(CLOCK_MONOTONIC)
 * Windows: using QueryPerformanceCounter()

Todo:
 * enable clock_gettime() on more platforms
 * add Mac OSX version, e.g. using mach_absolute_time + mach_timebase_info

Signed-off-by: Karsten Blees <blees@dcon.de>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

1 files changed, 82 insertions, 0 deletions

diff --git a/trace.c b/trace.c
index f013958d0a..b9d7272318 100644
--- a/trace.c
+++ b/trace.c
@@ -275,3 +275,85 @@ int trace_want(struct trace_key *key)
 {
 	return !!get_trace_fd(key);
 }
+
+#ifdef HAVE_CLOCK_GETTIME
+
+static inline uint64_t highres_nanos(void)
+{
+	struct timespec ts;
+	if (clock_gettime(CLOCK_MONOTONIC, &ts))
+		return 0;
+	return (uint64_t) ts.tv_sec * 1000000000 + ts.tv_nsec;
+}
+
+#elif defined (GIT_WINDOWS_NATIVE)
+
+static inline uint64_t highres_nanos(void)
+{
+	static uint64_t high_ns, scaled_low_ns;
+	static int scale;
+	LARGE_INTEGER cnt;
+
+	if (!scale) {
+		if (!QueryPerformanceFrequency(&cnt))
+			return 0;
+
+		/* high_ns = number of ns per cnt.HighPart */
+		high_ns = (1000000000LL << 32) / (uint64_t) cnt.QuadPart;
+
+		/*
+		 * Number of ns per cnt.LowPart is 10^9 / frequency (or
+		 * high_ns >> 32). For maximum precision, we scale this factor
+		 * so that it just fits within 32 bit (i.e. won't overflow if
+		 * multiplied with cnt.LowPart).
+		 */
+		scaled_low_ns = high_ns;
+		scale = 32;
+		while (scaled_low_ns >= 0x100000000LL) {
+			scaled_low_ns >>= 1;
+			scale--;
+		}
+	}
+
+	/* if QPF worked on initialization, we expect QPC to work as well */
+	QueryPerformanceCounter(&cnt);
+
+	return (high_ns * cnt.HighPart) +
+	       ((scaled_low_ns * cnt.LowPart) >> scale);
+}
+
+#else
+# define highres_nanos() 0
+#endif
+
+static inline uint64_t gettimeofday_nanos(void)
+{
+	struct timeval tv;
+	gettimeofday(&tv, NULL);
+	return (uint64_t) tv.tv_sec * 1000000000 + tv.tv_usec * 1000;
+}
+
+/*
+ * Returns nanoseconds since the epoch (01/01/1970), for performance tracing
+ * (i.e. favoring high precision over wall clock time accuracy).
+ */
+inline uint64_t getnanotime(void)
+{
+	static uint64_t offset;
+	if (offset > 1) {
+		/* initialization succeeded, return offset + high res time */
+		return offset + highres_nanos();
+	} else if (offset == 1) {
+		/* initialization failed, fall back to gettimeofday */
+		return gettimeofday_nanos();
+	} else {
+		/* initialize offset if high resolution timer works */
+		uint64_t now = gettimeofday_nanos();
+		uint64_t highres = highres_nanos();
+		if (highres)
+			offset = now - highres;
+		else
+			offset = 1;
+		return now;
+	}
+}