/**
* Licensed to the Apache Software Foundation (ASF) under one
* or more contributor license agreements. See the NOTICE file
* distributed with this work for additional information
* regarding copyright ownership. The ASF licenses this file
* to you under the Apache License, Version 2.0 (the
* "License"); you may not use this file except in compliance
* with the License. You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
package org.apache.hadoop.hdfs.server.namenode;
import java.io.IOException;
import java.util.ArrayList;
import java.util.Collection;
import java.util.Collections;
import java.util.Comparator;
import java.util.HashMap;
import java.util.HashSet;
import java.util.LinkedHashMap;
import java.util.List;
import java.util.Map;
import java.util.Set;
import org.apache.commons.logging.Log;
import org.apache.commons.logging.LogFactory;
import org.apache.hadoop.conf.Configuration;
import org.apache.hadoop.fs.Path;
import org.apache.hadoop.hdfs.protocol.Block;
import org.apache.hadoop.hdfs.protocol.DatanodeInfo;
import org.apache.hadoop.hdfs.protocol.LocatedBlock;
import org.apache.hadoop.hdfs.server.namenode.BlockPlacementPolicyDefault;
import org.apache.hadoop.net.DNSToSwitchMapping;
import org.apache.hadoop.net.NetworkTopology;
import org.apache.hadoop.net.Node;
import org.apache.hadoop.raid.RaidNode;
import org.apache.hadoop.raid.Codec;
import org.apache.hadoop.util.HostsFileReader;
import org.apache.hadoop.util.StringUtils;
/**
* This BlockPlacementPolicy uses a simple heuristic, random placement of
* the replicas of a newly-created block, for the purpose of spreading out the
* group of blocks which used by RAID for recovering each other.
* This is important for the availability of the blocks.
*
* Replication of an existing block continues to use the default placement
* policy.
*
* This simple block placement policy does not guarantee that
* blocks on the RAID stripe are on different nodes. However, BlockMonitor
* will periodically scans the raided files and will fix the placement
* if it detects violation.
*
* This class can be used by multiple threads. It has to be thread safe.
*/
public class BlockPlacementPolicyRaid extends BlockPlacementPolicyDefault {
public static final Log LOG =
LogFactory.getLog(BlockPlacementPolicyRaid.class);
Configuration conf;
private FSNamesystem namesystem = null;
private CachedLocatedBlocks cachedLocatedBlocks;
private CachedFullPathNames cachedFullPathNames;
/** {@inheritDoc} */
@Override
public void initialize(Configuration conf, FSClusterStats stats,
NetworkTopology clusterMap, HostsFileReader hostsReader,
DNSToSwitchMapping dnsToSwitchMapping, FSNamesystem namesystem) {
super.initialize(conf, stats, clusterMap,
hostsReader, dnsToSwitchMapping, namesystem);
this.conf = conf;
this.namesystem = namesystem;
this.cachedLocatedBlocks = new CachedLocatedBlocks(conf, namesystem);
this.cachedFullPathNames = new CachedFullPathNames(conf, namesystem);
}
@Override
public DatanodeDescriptor[] chooseTarget(String srcPath, int numOfReplicas,
DatanodeDescriptor writer, List<DatanodeDescriptor> chosenNodes,
long blocksize) {
return chooseTarget(srcPath, numOfReplicas, writer, chosenNodes, null,
blocksize);
}
@Override
public DatanodeDescriptor[] chooseTarget(String srcPath, int numOfReplicas,
DatanodeDescriptor writer, List<DatanodeDescriptor> chosenNodes,
List<Node> exlcNodes, long blocksize) {
try {
FileInfo info = getFileInfo(srcPath);
if (LOG.isDebugEnabled()) {
LOG.debug("FileType:" + srcPath + " " + info.type.name());
}
if (info.type == FileType.NOT_RAID) {
return super.chooseTarget(
srcPath, numOfReplicas, writer, chosenNodes, exlcNodes, blocksize);
}
ArrayList<DatanodeDescriptor> results = new ArrayList<DatanodeDescriptor>();
HashMap<Node, Node> excludedNodes = new HashMap<Node, Node>();
if (exlcNodes != null) {
for (Node node: exlcNodes) {
excludedNodes.put(node, node);
}
}
for (Node node:chosenNodes) {
excludedNodes.put(node, node);
}
chooseRandom(numOfReplicas, "/", excludedNodes, blocksize,
1, results);
return results.toArray(new DatanodeDescriptor[results.size()]);
} catch (Exception e) {
FSNamesystem.LOG.debug(
"Error happend when choosing datanode to write:" +
StringUtils.stringifyException(e));
return super.chooseTarget(srcPath, numOfReplicas, writer,
chosenNodes, blocksize);
}
}
/** {@inheritDoc} */
@Override
public DatanodeDescriptor chooseReplicaToDelete(FSInodeInfo inode,
Block block, short replicationFactor,
Collection<DatanodeDescriptor> first,
Collection<DatanodeDescriptor> second) {
DatanodeDescriptor chosenNode = null;
try {
String path = cachedFullPathNames.get(inode);
FileInfo info = getFileInfo(path);
if (info.type == FileType.NOT_RAID) {
return super.chooseReplicaToDelete(
inode, block, replicationFactor, first, second);
}
List<LocatedBlock> companionBlocks =
getCompanionBlocks(path, info, block);
if (companionBlocks == null || companionBlocks.size() == 0) {
// Use the default method if it is not a valid raided or parity file
return super.chooseReplicaToDelete(
inode, block, replicationFactor, first, second);
}
// Delete from the first collection first
// This ensures the number of unique rack of this block is not reduced
Collection<DatanodeDescriptor> all = new HashSet<DatanodeDescriptor>();
all.addAll(first);
all.addAll(second);
chosenNode = chooseReplicaToDelete(companionBlocks, all);
if (chosenNode != null) {
return chosenNode;
}
return super.chooseReplicaToDelete(
inode, block, replicationFactor, first, second);
} catch (Exception e) {
LOG.debug("Failed to choose the correct replica to delete", e);
return super.chooseReplicaToDelete(
inode, block, replicationFactor, first, second);
}
}
private DatanodeDescriptor chooseReplicaToDelete(
Collection<LocatedBlock> companionBlocks,
Collection<DatanodeDescriptor> dataNodes) throws IOException {
if (dataNodes.isEmpty()) {
return null;
}
// Count the number of replicas on each node and rack
final Map<String, Integer> nodeCompanionBlockCount =
countCompanionBlocks(companionBlocks, false);
final Map<String, Integer> rackCompanionBlockCount =
countCompanionBlocks(companionBlocks, true);
NodeComparator comparator =
new NodeComparator(nodeCompanionBlockCount, rackCompanionBlockCount);
return Collections.max(dataNodes, comparator);
}
/**
* Count how many companion blocks are on each datanode or the each rack
* @param companionBlocks a collection of all the companion blocks
* @param doRackCount count the companion blocks on the racks of datanodes
* @param result the map from node name to the number of companion blocks
*/
static Map<String, Integer> countCompanionBlocks(
Collection<LocatedBlock> companionBlocks, boolean doRackCount) {
Map<String, Integer> result = new HashMap<String, Integer>();
for (LocatedBlock block : companionBlocks) {
for (DatanodeInfo d : block.getLocations()) {
String name = doRackCount ? d.getParent().getName() : d.getName();
if (result.containsKey(name)) {
int count = result.get(name) + 1;
result.put(name, count);
} else {
result.put(name, 1);
}
}
}
return result;
}
/**
* Compares the datanodes based on the number of companion blocks on the same
* node and rack. If even, compare the remaining space on the datanodes.
*/
class NodeComparator implements Comparator<DatanodeDescriptor> {
private Map<String, Integer> nodeBlockCount;
private Map<String, Integer> rackBlockCount;
private NodeComparator(Map<String, Integer> nodeBlockCount,
Map<String, Integer> rackBlockCount) {
this.nodeBlockCount = nodeBlockCount;
this.rackBlockCount = rackBlockCount;
}
@Override
public int compare(DatanodeDescriptor d1, DatanodeDescriptor d2) {
int res = compareBlockCount(d1, d2, nodeBlockCount);
if (res != 0) {
return res;
}
res = compareBlockCount(d1.getParent(), d2.getParent(), rackBlockCount);
if (res != 0) {
return res;
}
if (d1.getRemaining() > d2.getRemaining()) {
return -1;
}
if (d1.getRemaining() < d2.getRemaining()) {
return 1;
}
return 0;
}
private int compareBlockCount(Node node1, Node node2,
Map<String, Integer> blockCount) {
Integer count1 = blockCount.get(node1.getName());
Integer count2 = blockCount.get(node2.getName());
count1 = count1 == null ? 0 : count1;
count2 = count2 == null ? 0 : count2;
if (count1 > count2) {
return 1;
}
if (count1 < count2) {
return -1;
}
return 0;
}
}
/**
* Obtain the companion blocks of the give block
* Companion blocks are defined as the blocks that can help recover each
* others by using raid decoder.
* @param path The path of the file contains the block
* @param info The info of this file
* @param block The given block
* null if it is the block which is currently being written to
* @return the block locations of companion blocks
*/
List<LocatedBlock> getCompanionBlocks(String path, FileInfo info, Block block)
throws IOException {
Codec codec = info.codec;
switch (info.type) {
case NOT_RAID:
return Collections.emptyList();
case HAR_TEMP_PARITY:
return getCompanionBlocksForHarParityBlock(
path, codec.parityLength, block);
case TEMP_PARITY:
return getCompanionBlocksForParityBlock(
getSourceFile(path, codec.tmpParityDirectory),
path, codec.parityLength, codec.stripeLength, block,
codec.isDirRaid);
case PARITY:
return getCompanionBlocksForParityBlock(
getSourceFile(path, codec.parityDirectory),
path, codec.parityLength, codec.stripeLength, block,
codec.isDirRaid);
case SOURCE:
return getCompanionBlocksForSourceBlock(
path,
getParityFile(codec, path),
codec.parityLength, codec.stripeLength, block,
codec.isDirRaid);
}
return Collections.emptyList();
}
private List<LocatedBlock> getCompanionBlocksForHarParityBlock(
String parity, int parityLength, Block block)
throws IOException {
int blockIndex = getBlockIndex(parity, block);
// consider only parity file in this case because source file block
// location is not easy to obtain
List<LocatedBlock> parityBlocks = cachedLocatedBlocks.get(parity);
List<LocatedBlock> result = new ArrayList<LocatedBlock>();
int start = Math.max(0, blockIndex - parityLength + 1);
int end = Math.min(parityBlocks.size(), blockIndex + parityLength);
result.addAll(parityBlocks.subList(start, end));
return result;
}
/**
* Check if we will do the directory raid on the file.
*
* @param fileBlocks, all the blocks of the file.
* @return
*/
private boolean canDoDirectoryRaid(List<LocatedBlock> fileBlocks) {
long minFileSize = conf.getLong(RaidNode.MINIMUM_RAIDABLE_FILESIZE_KEY,
RaidNode.MINIMUM_RAIDABLE_FILESIZE);
long fileSize = 0;
for (LocatedBlock block : fileBlocks) {
fileSize += block.getBlock().getNumBytes();
if (fileSize >= minFileSize) {
return true;
}
}
return false;
}
private List<LocatedBlock> getCompanionBlocksForParityBlock(
String src, String parity, int parityLength, int stripeLength,
Block block, boolean isDirRaid)
throws IOException {
int blockIndex = getBlockIndex(parity, block);
List<LocatedBlock> result = new ArrayList<LocatedBlock>();
List<LocatedBlock> parityBlocks = cachedLocatedBlocks.get(parity);
int stripeIndex = blockIndex / parityLength;
int parityStart = stripeIndex * parityLength;
int parityEnd = Math.min(parityStart + parityLength,
parityBlocks.size());
// for parity, always consider the neighbor blocks as companion blocks
if (parityStart < parityBlocks.size()) {
result.addAll(parityBlocks.subList(parityStart, parityEnd));
}
if (src == null) {
return result;
}
// get the source blocks.
List<LocatedBlock> sourceBlocks;
int sourceStart = stripeIndex * stripeLength;
int sourceEnd = sourceStart + stripeLength;
if (!isDirRaid) {
sourceBlocks = cachedLocatedBlocks.get(src);
} else {
sourceBlocks = new ArrayList<LocatedBlock>();
INode inode = namesystem.dir.getInode(src);
INodeDirectory srcNode;
if (inode.isDirectory()) {
srcNode = (INodeDirectory) inode;
} else {
throw new IOException(
"The source should be a directory in Dir-Raiding: " + src);
}
boolean found = false;
// look for the stripe
for (INode child : srcNode.getChildren()) {
if (child.isDirectory()) {
throw new IOException("The source is not a leaf directory: " + src
+ ", contains a subdirectory: " + child.getLocalName());
}
List<LocatedBlock> childBlocks =
cachedLocatedBlocks.get(src + "/" + child.getLocalName());
// check if we will do dir-raid on this file
if (!canDoDirectoryRaid(childBlocks)) {
continue;
}
if (childBlocks.size() < sourceStart && !found) {
sourceStart -= childBlocks.size();
sourceEnd -= childBlocks.size();
continue;
} else {
found = true;
sourceBlocks.addAll(childBlocks);
if (sourceEnd <= sourceBlocks.size()) {
break;
}
}
}
}
sourceEnd = Math.min(sourceEnd,
sourceBlocks.size());
if (sourceStart < sourceBlocks.size()) {
result.addAll(sourceBlocks.subList(sourceStart, sourceEnd));
}
return result;
}
private List<LocatedBlock> getCompanionBlocksForSourceBlock(
String src, String parity, int parityLength, int stripeLength,
Block block, boolean isDirRaid)
throws IOException {
int blockIndex = getBlockIndex(src, block);
List<LocatedBlock> result = new ArrayList<LocatedBlock>();
List<LocatedBlock> sourceBlocks;
int stripeIndex = 0;
int sourceStart = 0;
int sourceEnd = 0;
if (!isDirRaid) {
sourceBlocks = cachedLocatedBlocks.get(src);
stripeIndex = blockIndex / stripeLength;
sourceStart = stripeIndex * stripeLength;
sourceEnd = Math.min(sourceStart + stripeLength, sourceBlocks.size());
} else {
// cache the candidate blocks.
LocatedBlock[] tmpStripe = new LocatedBlock[stripeLength];
int curIdx = 0;
boolean found = false;
sourceBlocks = new ArrayList<LocatedBlock>();
INodeDirectory srcNode = namesystem.dir.getInode(src).getParent();
String srcPath = new Path(src).getParent().toString();
// look for the stripe
for (INode child : srcNode.getChildren()) {
if (child.isDirectory()) {
throw new IOException("The raided-directory is not a leaf directory: "
+ srcPath +
", contains a subdirectory: " + child.getLocalName());
}
String childName = srcPath + "/" + child.getLocalName();
List<LocatedBlock> childBlocks = cachedLocatedBlocks.get(childName);
// check if we will do dir-raid on this file
if (!canDoDirectoryRaid(childBlocks)) {
continue;
}
if (found) {
if (sourceEnd <= sourceBlocks.size()) {
break;
}
sourceBlocks.addAll(childBlocks);
} else {
int childBlockSize = childBlocks.size();
/**
* If we find the target file, we will addAll the
* cached blocks and the child blocks.
* And update the metrics like stripeIndex, sourceStart and sourceEnd.
*
*/
if (childName.equals(src)) {
found = true;
for (int i=0; i<curIdx; i++) {
sourceBlocks.add(tmpStripe[i]);
}
sourceBlocks.addAll(childBlocks);
blockIndex += curIdx;
stripeIndex += blockIndex / stripeLength;
sourceStart = (blockIndex / stripeLength) * stripeLength;
sourceEnd = sourceStart + stripeLength;
} else {
/**
* If not find the target file, we will keep the current stripe
* in the temp stripe cache.
*/
/**
* the childBlockSize is small, and we can fill them into
* current temp stripe cache.
*/
if (curIdx + childBlockSize < stripeLength) {
for (int i=0; i<childBlockSize; i++) {
tmpStripe[curIdx++] = childBlocks.get(i);
}
} else {
/**
* The childBlockSize is not small, We need to calculate
* the place in the stripe cache, and copy the current stripe
* into the temp stripe cache.
*/
stripeIndex += (curIdx + childBlockSize) / stripeLength;
int childStart = ((curIdx + childBlockSize) / stripeLength)
* stripeLength - curIdx;
curIdx = 0;
for (; childStart<childBlockSize;) {
tmpStripe[curIdx++] = childBlocks.get(childStart++);
}
curIdx %= stripeLength;
}
}
}
}
sourceEnd = Math.min(sourceEnd, sourceBlocks.size());
}
if (sourceStart < sourceBlocks.size()) {
result.addAll(sourceBlocks.subList(sourceStart, sourceEnd));
}
if (parity == null) {
return result;
}
// get the parity blocks.
List<LocatedBlock> parityBlocks = cachedLocatedBlocks.get(parity);
int parityStart = stripeIndex * parityLength;
int parityEnd = Math.min(parityStart + parityLength,
parityBlocks.size());
if (parityStart < parityBlocks.size()) {
result.addAll(parityBlocks.subList(parityStart, parityEnd));
}
return result;
}
private int getBlockIndex(String file, Block block) throws IOException {
List<LocatedBlock> blocks = cachedLocatedBlocks.get(file);
// null indicates that this block is currently added. Return size()
// as the index in this case
if (block == null) {
return blocks.size();
}
for (int i = 0; i < blocks.size(); i++) {
if (blocks.get(i).getBlock().equals(block)) {
return i;
}
}
throw new IOException("Cannot locate " + block + " in file " + file);
}
/**
* Cache results for FSInodeInfo.getFullPathName()
*/
static class CachedFullPathNames {
private Cache<INodeWithHashCode, String> cacheInternal;
private FSNamesystem namesystem;
CachedFullPathNames(final Configuration conf, final FSNamesystem namesystem) {
this.namesystem = namesystem;
this.cacheInternal = new Cache<INodeWithHashCode, String>(conf) {
@Override
public String getDirectly(INodeWithHashCode inode) throws IOException {
namesystem.readLock();
try {
return inode.getFullPathName();
} finally {
namesystem.readUnlock();
}
}
};
}
private static class INodeWithHashCode {
FSInodeInfo inode;
INodeWithHashCode(FSInodeInfo inode) {
this.inode = inode;
}
@Override
public boolean equals(Object obj) {
return inode == obj;
}
@Override
public int hashCode() {
return System.identityHashCode(inode);
}
String getFullPathName() {
return inode.getFullPathName();
}
}
public String get(FSInodeInfo inode) throws IOException {
return cacheInternal.get(new INodeWithHashCode(inode));
}
}
/**
* Cache results for FSNamesystem.getBlockLocations()
*/
static class CachedLocatedBlocks extends Cache<String, List<LocatedBlock>> {
private FSNamesystem namesystem;
CachedLocatedBlocks(Configuration conf, FSNamesystem namesystem) {
super(conf);
this.namesystem = namesystem;
}
@Override
public List<LocatedBlock> getDirectly(String file) throws IOException {
INodeFile inode = namesystem.dir.getFileINode(file);
// Note that the list is generated. It is not the internal data of inode.
List<LocatedBlock> result = inode == null ?
new ArrayList<LocatedBlock>() :
namesystem.getBlockLocationsInternal(inode, 0, Long.MAX_VALUE,
Integer.MAX_VALUE).getLocatedBlocks();
if (result == null) {
return Collections.emptyList();
}
return Collections.unmodifiableList(result);
}
}
private static abstract class Cache<K, V> {
private Map<K, ValueWithTime> cache;
final private long cacheTimeout;
final private int maxEntries;
// The timeout is long but the consequence of stale value is not serious
Cache(Configuration conf) {
this.cacheTimeout =
conf.getLong("raid.blockplacement.cache.timeout", 5000L); // 5 seconds
this.maxEntries =
conf.getInt("raid.blockplacement.cache.size", 1000); // 1000 entries
Map<K, ValueWithTime> map = new LinkedHashMap<K, ValueWithTime>(
maxEntries, 0.75f, true) {
private static final long serialVersionUID = 1L;
@Override
protected boolean removeEldestEntry(
Map.Entry<K, ValueWithTime> eldest) {
return size() > maxEntries;
}
};
this.cache = Collections.synchronizedMap(map);
}
// Note that this method may hold FSNamesystem.readLock() and it may
// be called inside FSNamesystem.writeLock(). If we make this method
// synchronized, it will deadlock.
abstract protected V getDirectly(K key) throws IOException;
public V get(K key) throws IOException {
// The method is not synchronized so we may get some stale value here but
// it's OK.
ValueWithTime result = cache.get(key);
long now = System.currentTimeMillis();
if (result != null &&
now - result.cachedTime < cacheTimeout) {
return result.value;
}
result = new ValueWithTime();
result.value = getDirectly(key);
result.cachedTime = now;
cache.put(key, result);
return result.value;
}
private class ValueWithTime {
V value = null;
long cachedTime = 0L;
}
}
/**
* Get path for the corresponding source file for a valid parity
* file. Returns null if it does not exists
* @param parity the toUri path of the parity file
* @return the toUri path of the source file
*/
String getSourceFile(String parity, String prefix) throws IOException {
if (isHarFile(parity)) {
return null;
}
// remove the prefix
String src = parity.substring(prefix.length());
if (namesystem.dir.getInode(src) == null) {
return null;
}
return src;
}
/**
* Get path for the parity file. Returns null if it does not exists
* @param codec the codec of the parity file.
* @return the toUri path of the parity file
*/
private String getParityFile(Codec codec, String src)
throws IOException {
String parity;
if (codec.isDirRaid) {
String parent = new Path(src).getParent().toString();
parity = codec.parityDirectory + parent;
} else {
parity = codec.parityDirectory + src;
}
if (namesystem.dir.getInode(parity) == null) {
return null;
}
return parity;
}
private boolean isHarFile(String path) {
return path.lastIndexOf(RaidNode.HAR_SUFFIX) != -1;
}
class FileInfo {
FileInfo(FileType type, Codec codec) {
this.type = type;
this.codec = codec;
}
final FileType type;
final Codec codec;
}
enum FileType {
NOT_RAID,
HAR_TEMP_PARITY,
TEMP_PARITY,
PARITY,
SOURCE,
}
/**
* Return raid information about a file, for example
* if this file is the source file, parity file, or not raid
*
* @param path file name
* @return raid information
* @throws IOException
*/
protected FileInfo getFileInfo(String path) throws IOException {
for (Codec c : Codec.getCodecs()) {
if (path.startsWith(c.tmpHarDirectory + Path.SEPARATOR)) {
return new FileInfo(FileType.HAR_TEMP_PARITY, c);
}
if (path.startsWith(c.tmpParityDirectory + Path.SEPARATOR)) {
return new FileInfo(FileType.TEMP_PARITY, c);
}
if (path.startsWith(c.parityDirectory + Path.SEPARATOR)) {
return new FileInfo(FileType.PARITY, c);
}
String parity = getParityFile(c, path);
if (parity != null) {
return new FileInfo(FileType.SOURCE, c);
}
}
return new FileInfo(FileType.NOT_RAID, null);
}
}