From: Wouter Wijngaards <wouter@nlnetlabs.nl>
Date: Tue, 19 Oct 2010 14:53:29 +0000 (+0000)
Subject: Fix for request list growth.
X-Git-Tag: release-1.4.7rc1~27
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e238b47b0accedfaaa67491ec13b491bc8c3eab2;p=thirdparty%2Funbound.git

Fix for request list growth.


git-svn-id: file:///svn/unbound/trunk@2298 be551aaa-1e26-0410-a405-d3ace91eadb9
---

diff --git a/doc/Changelog b/doc/Changelog
index 7892a964e..0c88f076f 100644
--- a/doc/Changelog
+++ b/doc/Changelog
@@ -1,3 +1,11 @@
+19 October 2010: Wouter
+	- Fix for request list growth, if a server has long timeout but the
+	  lost counter is low, then its effective rtt is the one without
+	  exponential backoff applied.  Because the backoff is not working.
+	  The lost counter can then increase and the server is blacklisted,
+	  or the lost counter does not increase and the server is working
+	  for some queries.
+
 18 October 2010: Wouter
 	- iana portlist updated.
 
diff --git a/iterator/iter_utils.c b/iterator/iter_utils.c
index 1ac7aa265..ca1781a1b 100644
--- a/iterator/iter_utils.c
+++ b/iterator/iter_utils.c
@@ -208,11 +208,6 @@ iter_filter_unsuitable(struct iter_env* iter_env, struct module_env* env,
 			return -1; /* server is lame */
 		else if(rtt >= USEFUL_SERVER_TOP_TIMEOUT && 
 			lost >= USEFUL_SERVER_MAX_LOST) {
-			/* keep trying slowly, 1% of the time, because
-			 * this can be due to weird firewalls.  This number
-			 * does not have to be securely random. */
-			if(ub_random(env->rnd) % 100 == 0)
-				return USEFUL_SERVER_TOP_TIMEOUT+1;
 			/* server is unresponsive */
 			return USEFUL_SERVER_TOP_TIMEOUT;
 		}
diff --git a/services/cache/infra.c b/services/cache/infra.c
index 219a67a6b..a43c22b0b 100644
--- a/services/cache/infra.c
+++ b/services/cache/infra.c
@@ -47,6 +47,7 @@
 #include "util/log.h"
 #include "util/net_help.h"
 #include "util/config_file.h"
+#include "iterator/iterator.h"
 
 size_t 
 infra_host_sizefunc(void* k, void* ATTR_UNUSED(d))
@@ -267,6 +268,14 @@ infra_host(struct infra_cache* infra, struct sockaddr_storage* addr,
 	/* use existing entry */
 	data = (struct infra_host_data*)e->data;
 	*to = rtt_timeout(&data->rtt);
+	if(*to >= USEFUL_SERVER_TOP_TIMEOUT &&
+		data->num_timeouts < USEFUL_SERVER_MAX_LOST)
+		/* use smaller timeout, backoff does not work
+		 * The server seems to still reply but sporadically.
+		 * Perhaps it has rate-limited the traffic, or it
+		 * drops particular queries (AAAA).  ignore timeouts,
+		 * but we use an expanded variance of 6x. */
+		*to = data->rtt.srtt + 6*data->rtt.rttvar;
 	*edns_vs = data->edns_version;
 	*edns_lame_known = data->edns_lame_known;
 	lock_rw_unlock(&e->lock);