Fix: handle negative clock offset

In the unlikely situation where a system sets its hardware clock
(CLOCK_REALTIME) to 0 (Epoch) after boot, the difference

  monotonic - realtime

becomes negative.

Fixup this situation by returning a 0 offset in this case.

This ensures that trace viewer implementations (e.g. babeltrace) which
currently cannot handle the negative offset (known bug) still work with
the generated traces.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>

diff --git a/lttng-events.c b/lttng-events.c
index f0f5d0f80dbe340bf1e4663681ff621acd25f504..bd6f44d977dba95e36c769f828b60e4fb9502c1f 100644 (file)
--- a/lttng-events.c
+++ b/lttng-events.c
@@ -2023,12 +2023,15 @@ int _lttng_event_header_declare(struct lttng_session *session)
  * taken at start of trace.
  * Yes, this is only an approximation. Yes, we can (and will) do better
  * in future versions.
+ * Return 0 if offset is negative. It may happen if the system sets
+ * the REALTIME clock to 0 after boot.
  */
 static
 uint64_t measure_clock_offset(void)
 {
-       uint64_t offset, monotonic[2], realtime;
+       uint64_t monotonic_avg, monotonic[2], realtime;
        uint64_t tcf = trace_clock_freq();
+       int64_t offset;
        struct timespec rts = { 0, 0 };
        unsigned long flags;
 
@@ -2039,7 +2042,7 @@ uint64_t measure_clock_offset(void)
        monotonic[1] = trace_clock_read64();
        local_irq_restore(flags);
 
-       offset = (monotonic[0] + monotonic[1]) >> 1;
+       monotonic_avg = (monotonic[0] + monotonic[1]) >> 1;
        realtime = (uint64_t) rts.tv_sec * tcf;
        if (tcf == NSEC_PER_SEC) {
                realtime += rts.tv_nsec;
@@ -2049,7 +2052,9 @@ uint64_t measure_clock_offset(void)
                do_div(n, NSEC_PER_SEC);
                realtime += n;
        }
-       offset = realtime - offset;
+       offset = (int64_t) realtime - monotonic_avg;
+       if (offset < 0)
+               return 0;
        return offset;
 }