HDFS-4937. ReplicationMonitor can infinite-loop in BlockPlacementPolicyDefault#chooseRandom(). Contributed by Kihwal Lee.

hadoop-hdfs-project/hadoop-hdfs/CHANGES.txt

@@ -2201,6 +2201,9 @@ Release 2.8.0 - UNRELEASED
     HDFS-9332. Fix Precondition failures from NameNodeEditLogRoller while
     saving namespace. (wang)
 
+    HDFS-4937. ReplicationMonitor can infinite-loop in
+    BlockPlacementPolicyDefault#chooseRandom() (kihwal)
+
 Release 2.7.2 - UNRELEASED
 
   INCOMPATIBLE CHANGES

hadoop-hdfs-project/hadoop-hdfs/src/main/java/org/apache/hadoop/hdfs/server/blockmanagement/BlockPlacementPolicyDefault.java

@@ -659,6 +659,7 @@ public class BlockPlacementPolicyDefault extends BlockPlacementPolicy {
 
     int numOfAvailableNodes = clusterMap.countNumOfAvailableNodes(
         scope, excludedNodes);
+    int refreshCounter = numOfAvailableNodes;
     StringBuilder builder = null;
     if (LOG.isDebugEnabled()) {
       builder = debugLoggingBuilder.get();
@@ -708,6 +709,17 @@ public class BlockPlacementPolicyDefault extends BlockPlacementPolicy {
         // If no candidate storage was found on this DN then set badTarget.
         badTarget = (storage == null);
       }
+      // Refresh the node count. If the live node count became smaller,
+      // but it is not reflected in this loop, it may loop forever in case
+      // the replicas/rack cannot be satisfied.
+      if (--refreshCounter == 0) {
+        refreshCounter = clusterMap.countNumOfAvailableNodes(scope,
+            excludedNodes);
+        // It has already gone through enough number of nodes.
+        if (refreshCounter <= excludedNodes.size()) {
+          break;
+        }
+      }
     }
       
     if (numOfReplicas>0) {