HDDS-11261. Fix old block info cache when ec reconstruction failed.

hadoop-hdds/client/src/main/java/org/apache/hadoop/ozone/client/io/ECBlockInputStreamProxy.java

@@ -52,11 +52,12 @@ public class ECBlockInputStreamProxy extends BlockExtendedInputStream {
   private final ECReplicationConfig repConfig;
   private final XceiverClientFactory xceiverClientFactory;
   private final Function<BlockID, BlockLocationInfo> refreshFunction;
-  private final BlockLocationInfo blockInfo;
+  private BlockLocationInfo blockInfo;
   private final ECBlockInputStreamFactory ecBlockInputStreamFactory;
 
   private BlockExtendedInputStream blockReader;
   private boolean reconstructionReader = false;
+  private boolean refreshedCache = false;
   private List<DatanodeDetails> failedLocations = new ArrayList<>();
   private boolean closed = false;
   private OzoneClientConfig config;
@@ -130,6 +131,17 @@ private void createBlockReader() {
         xceiverClientFactory, refreshFunction, config);
   }
 
+  private void reCreateBlockReaderWithFreshBlockInfo() {
+    XceiverClientManager.getXceiverClientMetrics()
+        .incECReconstructionTotal();
+    blockInfo = refreshFunction.apply(getBlockID());
+    refreshedCache = true;
+
+    blockReader = ecBlockInputStreamFactory.create(reconstructionReader,
+        failedLocations, repConfig, blockInfo,
+        xceiverClientFactory, refreshFunction, config);
+  }
+
   @Override
   public synchronized BlockID getBlockID() {
     return blockInfo.getBlockID();
@@ -167,12 +179,21 @@ public synchronized int read(ByteBuffer buf) throws IOException {
       }
     } catch (IOException e) {
       if (reconstructionReader) {
-        // If we get an error from the reconstruction reader, there
-        // is nothing left to try. It will re-try until it has insufficient
-        // locations internally, so if an error comes here, just re-throw it.
-        XceiverClientManager.getXceiverClientMetrics()
-            .incECReconstructionFailsTotal();
-        throw e;
+        if (!refreshedCache && e instanceof InsufficientLocationsException) {
+          // When get key with insufficient nodes amount OM cache filled insufficient amount of nodes.
+          // When nodes amount become enough OM have old cache.
+          // In that case every request for this key will be finished with "insufficient nodes error" because of
+          // old cache.
+          // Need to refresh cache to get actual information about nodes.
+          return reReadWithRecreateBlockReader(buf, lastPosition, totalRead);
+        } else {
+          // If we get an error from the reconstruction reader with actual cache, there
+          // is nothing left to try. It will re-try until it has insufficient
+          // locations internally, so if an error comes here, just re-throw it.
+          XceiverClientManager.getXceiverClientMetrics()
+              .incECReconstructionFailsTotal();
+          throw e;
+        }
       }
       if (e instanceof BadDataLocationException) {
         String message = "Failing over to reconstruction read due" +
@@ -195,6 +216,18 @@ public synchronized int read(ByteBuffer buf) throws IOException {
     return totalRead;
   }
 
+  private int reReadWithRecreateBlockReader(ByteBuffer buf, long lastPosition, int totalRead) throws IOException {
+    blockReader.close();
+    reCreateBlockReaderWithFreshBlockInfo();
+
+    if (lastPosition != 0) {
+      blockReader.seek(lastPosition);
+    }
+    buf.reset();
+    totalRead += read(buf);
+    return totalRead;
+  }
+
   private synchronized void failoverToReconstructionRead(
       List<DatanodeDetails> badLocations, long lastPosition)
       throws IOException {