From: Vladimír Čunát <vladimir.cunat@nic.cz>
Date: Tue, 29 Dec 2020 08:28:16 +0000 (+0100)
Subject: lib/selection: be more careful around rtt_state.dead_since
X-Git-Tag: v5.3.0~30^2~4
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=846d576b7f2e560baf038b9cfbafdc4490bea45d;p=thirdparty%2Fknot-resolver.git

lib/selection: be more careful around rtt_state.dead_since

It's all because the timestamp that we're using isn't (guaranteed to be)
meaningful across reboots or different machines, whereas our cache even
persists by default.
---

diff --git a/lib/selection.c b/lib/selection.c
index 6dad1a296..0f88e15c8 100644
--- a/lib/selection.c
+++ b/lib/selection.c
@@ -179,8 +179,20 @@ static struct rtt_state calc_rtt_state(struct rtt_state old, unsigned new_rtt)
 static void invalidate_dead_upstream(struct address_state *state,
 				     unsigned int retry_timeout)
 {
-	if (kr_now() - state->rtt_state.dead_since < retry_timeout) {
-		state->generation = -1;
+	struct rtt_state *rs = &state->rtt_state;
+	if (rs->consecutive_timeouts >= KR_NS_TIMEOUT_ROW_DEAD) {
+		uint64_t now = kr_now();
+		if (now < rs->dead_since) {
+			// broken continuity of timestamp (reboot, different machine, etc.)
+			*rs = default_rtt_state;
+		} else if (now < rs->dead_since + retry_timeout) {
+			// period when we don't want to use the address
+			state->generation = -1;
+		} else {
+			assert(now >= rs->dead_since + retry_timeout);
+			// we allow to retry the server now
+			// TODO: perhaps tweak *rs?
+		}
 	}
 }