modules/stats: further improve latency measurements

libuv's updating the stamp just once per event loop
might be too coarse, as processing multiple packets
(e.g. DNSSEC validation) may take milliseconds together.

Of course we still don't measure e.g. the time when incoming
client requests stay in OS buffers.

(cherry picked from commit 5e6a02b)

Author: vcunat

NEWS

@@ -10,6 +10,10 @@ Improvements
 - tests: disable problematic config.http test (#925, !1678)
 - validator: accept a confusing NODATA proof with insecure delegation (!1678)
 
+Bugfixes
+--------
+- stats: request latency was very incorrect in some cases (!1678)
+
 
 Knot Resolver 5.7.4 (2024-07-23)
 ================================

lib/rplan.c

@@ -157,6 +157,8 @@ static struct kr_query *kr_rplan_push_query(struct kr_rplan *rplan,
 	qry->request = rplan->request;
 
 	gettimeofday(&qry->timestamp, NULL);
+	if (!parent)  // start of kr_request; let's make the stamp more precise
+		uv_update_time(uv_default_loop());
 	qry->timestamp_mono = kr_now();
 	qry->creation_time_mono = parent ? parent->creation_time_mono : qry->timestamp_mono;
 	kr_zonecut_init(&qry->zone_cut, (const uint8_t *)"", rplan->pool);

modules/stats/stats.c

@@ -248,8 +248,14 @@ static int collect(kr_layer_t *ctx)
 	collect_answer(data, param->answer);
 	/* Count cached and unresolved */
 	if (rplan->resolved.len > 0) {
-		/* Histogram of answer latency. */
+		/* Histogram of answer latency.
+		 *
+		 * We update the notion of time.  Once per .finish isn't that expensive.
+		 * defer_* also updates this if active, but not in ideal moment for stats.
+		 */
+		uv_update_time(uv_default_loop());
 		uint64_t elapsed = kr_now() - rplan->initial->creation_time_mono;
+
 		stat_const_add(data, metric_answer_sum_ms, elapsed);
 		if (elapsed <= 1) {
 			stat_const_add(data, metric_answer_1ms, 1);