From: Pieter Lexis <pieter.lexis@powerdns.com>
Date: Fri, 27 Jan 2017 15:57:41 +0000 (+0100)
Subject: Auth: Incremental backoff for failed slave checks
X-Git-Tag: rec-4.1.0-alpha1~260^2
X-Git-Url: https://granicus.if.org/sourcecode?a=commitdiff_plain;h=b239dfbac44fd35053b5760e64686445e8210569;p=pdns

Auth: Incremental backoff for failed slave checks

Closes #349
Closes #602
---

diff --git a/docs/markdown/authoritative/modes-of-operation.md b/docs/markdown/authoritative/modes-of-operation.md
index f48e37afa..5a9a88136 100644
--- a/docs/markdown/authoritative/modes-of-operation.md
+++ b/docs/markdown/authoritative/modes-of-operation.md
@@ -65,6 +65,10 @@ is higher, the domain is retrieved and inserted into the database. In any case,
 after the check the domain is declared 'fresh', and will only be checked again
 after '**refresh**' seconds have passed.
 
+When the freshness of a domain cannot be checked, e.g. because the master is offline, PowerDNS will retry the domain after [`slave-cycle-interval`](settings.md#slave-cycle-interval) seconds.
+Every time the domain fails it's freshness check, PowerDNS will hold back on checking the domain for `amount of failures * slave-cycle-interval` seconds, with a maximum of [`soa-retry-default`](settings.md#soa-retry-default) seconds between checks.
+With default settings, this means that PowerDNS will back off for 1, then 2, then 3 etc. minutes, to a maximum of 60 minutes between checks.
+
 **Warning**: Slave support is OFF by default, turn it on by adding [`slave`](settings.md#slave) to the configuration.
 **Note**: When running PowerDNS via the provided systemd service file, [`ProtectSystem`](http://www.freedesktop.org/software/systemd/man/systemd.exec.html#ProtectSystem=) is set to `full`, this means PowerDNS is unable to write to e.g. `/etc` and `/home`, possibly being unable to write AXFR's zones.
 
diff --git a/pdns/communicator.hh b/pdns/communicator.hh
index 8171ec4f6..d57c2ada4 100644
--- a/pdns/communicator.hh
+++ b/pdns/communicator.hh
@@ -217,6 +217,11 @@ private:
   bool d_masterschanged, d_slaveschanged;
   bool d_preventSelfNotification;
 
+  // Used to keep some state on domains that failed their freshness checks.
+  // uint64_t == counter of the number of failures (increased by 1 every consecutive slave-cycle-interval that the domain fails)
+  // time_t == wait at least until this time before attempting a new check
+  map<DNSName, pair<uint64_t, time_t> > d_failedSlaveRefresh;
+
   struct RemoveSentinel
   {
     explicit RemoveSentinel(const DNSName& dn, CommunicatorClass* cc) : d_dn(dn), d_cc(cc)
diff --git a/pdns/slavecommunicator.cc b/pdns/slavecommunicator.cc
index 0df99b075..b95b2c063 100644
--- a/pdns/slavecommunicator.cc
+++ b/pdns/slavecommunicator.cc
@@ -748,8 +748,13 @@ void CommunicatorClass::slaveRefresh(PacketHandler *P)
   {
     Lock l(&d_lock);
     domains_by_name_t& nameindex=boost::multi_index::get<IDTag>(d_suckdomains);
+    time_t now = time(0);
 
     for(DomainInfo& di :  rdomains) {
+      const auto failed = d_failedSlaveRefresh.find(di.zone);
+      if (failed != d_failedSlaveRefresh.end() && now < failed->second.second )
+        // If the domain has failed before and the time before the next check has not expired, skip this domain
+        continue;
       std::vector<std::string> localaddr;
       SuckRequest sr;
       sr.domain=di.zone;
@@ -828,6 +833,7 @@ void CommunicatorClass::slaveRefresh(PacketHandler *P)
   L<<Logger::Warning<<"Received serial number updates for "<<ssr.d_freshness.size()<<" zone"<<addS(ssr.d_freshness.size())<<", had "<<ifl.getTimeouts()<<" timeout"<<addS(ifl.getTimeouts())<<endl;
 
   typedef DomainNotificationInfo val_t;
+  time_t now = time(0);
   for(val_t& val :  sdomains) {
     DomainInfo& di(val.di);
     // might've come from the packethandler
@@ -836,8 +842,22 @@ void CommunicatorClass::slaveRefresh(PacketHandler *P)
         continue;
     }
 
-    if(!ssr.d_freshness.count(di.id)) // what does this mean? XXX
+    if(!ssr.d_freshness.count(di.id)) { // If we don't have an answer for the domain
+      uint64_t newCount = 1;
+      const auto failedEntry = d_failedSlaveRefresh.find(di.zone);
+      if (failedEntry != d_failedSlaveRefresh.end())
+        newCount = d_failedSlaveRefresh[di.zone].first + 1;
+      time_t nextCheck = now + std::min(newCount * d_tickinterval, (uint64_t)::arg().asNum("soa-retry-default"));
+      d_failedSlaveRefresh[di.zone] = {newCount, nextCheck};
+      if (newCount == 1 || newCount % 10 == 0)
+        L<<Logger::Warning<<"Unable to retrieve SOA for "<<di.zone<<", this was the "<<(newCount == 1 ? "first" : std::to_string(newCount) + "th")<<" time."<<endl;
       continue;
+    }
+
+    const auto wasFailedDomain = d_failedSlaveRefresh.find(di.zone);
+    if (wasFailedDomain != d_failedSlaveRefresh.end())
+      d_failedSlaveRefresh.erase(di.zone);
+
     uint32_t theirserial = ssr.d_freshness[di.id].theirSerial, ourserial = di.serial;
 
     if(rfc1982LessThan(theirserial, ourserial) && ourserial != 0) {