/** Resizes the internal byte buffer with a simple doubling policy, if needed. */
private final void growIfNeeded(int minimumDesired) {
if (buffer.b().remaining() < minimumDesired) {
// Compute the size of the new buffer
int newCapacity = buffer.b().capacity();
int newRemaining = newCapacity - buffer.b().position();
while (newRemaining < minimumDesired) {
newRemaining += newCapacity;
newCapacity *= 2;
}
// Allocate and copy
BBContainer next;
if (isDirect) {
next = DBBPool.allocateDirect(newCapacity);
} else {
next = DBBPool.wrapBB(ByteBuffer.allocate(newCapacity));
}
buffer.b().flip();
next.b().put(buffer.b());
assert next.b().remaining() == newRemaining;
buffer.discard();
buffer = next;
if (callback != null) callback.onBufferGrow(this);
assert (buffer.b().order() == ByteOrder.BIG_ENDIAN);
}
}
private void convertChunks() throws IOException, InterruptedException {
int lastNumCharacters = 1024 * 64;
while (!Thread.interrupted() && m_saveFile.hasMoreChunks()) {
if (m_availableBytes.get() > m_maxAvailableBytes) {
Thread.sleep(5);
continue;
}
BBContainer c = m_saveFile.getNextChunk();
if (c == null) {
return;
}
try {
final VoltTable vt = PrivateVoltTableFactory.createVoltTableFromBuffer(c.b(), true);
Pair<Integer, byte[]> p = VoltTableUtil.toCSV(vt, m_delimiter, null, lastNumCharacters);
lastNumCharacters = p.getFirst();
byte csvBytes[] = p.getSecond();
// should not insert empty byte[] if not last ConverterThread
if (csvBytes.length > 0) {
m_availableBytes.addAndGet(csvBytes.length);
m_available.offer(csvBytes);
}
} finally {
c.discard();
}
}
}
/**
* Get a ascii-string-safe version of the binary value using a hex encoding.
*
* @return A hex-encoded string value representing the serialized objects.
*/
public String getHexEncodedBytes() {
buffer.b().flip();
byte bytes[] = new byte[buffer.b().remaining()];
buffer.b().get(bytes);
String hex = Encoder.hexEncode(bytes);
buffer.discard();
return hex;
}
@Override
public void discard() {
if (m_hasMoreChunks == false) {
m_origin.discard();
} else {
m_buffers.add(this);
}
}
@Override
public ListenableFuture<?> write(Callable<BBContainer> tupleData, int tableId) {
try {
BBContainer container = tupleData.call();
if (container != null) {
container.discard();
}
} catch (Exception e) {
}
return null;
}
public void close() throws IOException {
if (m_chunkReaderThread != null) {
m_chunkReaderThread.interrupt();
try {
m_chunkReaderThread.join();
} catch (InterruptedException e) {
throw new IOException(e);
}
}
synchronized (this) {
while (!m_availableChunks.isEmpty()) {
m_availableChunks.poll().discard();
}
notifyAll();
}
/*
* Free buffers used to pull snapshot data in process
*/
BBContainer cont;
while ((cont = m_buffers.poll()) != null) {
cont.discard();
}
}
public static synchronized void pushDRBuffer(int partitionId, ByteBuffer buf) {
if (logDebug) {
System.out.println("Received DR buffer size " + buf.remaining());
AtomicLong haveOpenTransaction = haveOpenTransactionLocal.get();
buf.order(ByteOrder.LITTLE_ENDIAN);
// Magic header space for Java for implementing zero copy stuff
buf.position(8);
while (buf.hasRemaining()) {
int startPosition = buf.position();
byte version = buf.get();
int type = buf.get();
int checksum = 0;
if (version != 0) System.out.println("Remaining is " + buf.remaining());
switch (DRRecordType.valueOf(type)) {
case INSERT:
{
// Insert
if (haveOpenTransaction.get() == -1) {
System.out.println("Have insert but no open transaction");
System.exit(-1);
}
final long tableHandle = buf.getLong();
final int lengthPrefix = buf.getInt();
buf.position(buf.position() + lengthPrefix);
checksum = buf.getInt();
System.out.println(
"Version "
+ version
+ " type INSERT table handle "
+ tableHandle
+ " length "
+ lengthPrefix
+ " checksum "
+ checksum);
break;
}
case DELETE:
{
// Delete
if (haveOpenTransaction.get() == -1) {
System.out.println("Have insert but no open transaction");
System.exit(-1);
}
final long tableHandle = buf.getLong();
final int lengthPrefix = buf.getInt();
buf.position(buf.position() + lengthPrefix);
checksum = buf.getInt();
System.out.println(
"Version "
+ version
+ " type DELETE table handle "
+ tableHandle
+ " length "
+ lengthPrefix
+ " checksum "
+ checksum);
break;
}
case UPDATE:
// Update
// System.out.println("Version " + version + " type UPDATE " + checksum " + checksum);
break;
case BEGIN_TXN:
{
// Begin txn
final long txnId = buf.getLong();
final long spHandle = buf.getLong();
if (haveOpenTransaction.get() != -1) {
System.out.println(
"Have open transaction txnid "
+ txnId
+ " spHandle "
+ spHandle
+ " but already open transaction");
System.exit(-1);
}
haveOpenTransaction.set(spHandle);
checksum = buf.getInt();
System.out.println(
"Version "
+ version
+ " type BEGIN_TXN "
+ " txnid "
+ txnId
+ " spHandle "
+ spHandle
+ " checksum "
+ checksum);
break;
}
case END_TXN:
{
// End txn
final long spHandle = buf.getLong();
if (haveOpenTransaction.get() == -1) {
System.out.println(
"Have end transaction spHandle "
+ spHandle
+ " but no open transaction and its less then last committed "
+ lastCommittedSpHandle.get().get());
// checksum = buf.getInt();
// break;
System.exit(-1);
}
haveOpenTransaction.set(-1);
lastCommittedSpHandle.get().set(spHandle);
checksum = buf.getInt();
System.out.println(
"Version "
+ version
+ " type END_TXN "
+ " spHandle "
+ spHandle
+ " checksum "
+ checksum);
break;
}
}
int calculatedChecksum =
DBBPool.getBufferCRC32C(buf, startPosition, buf.position() - startPosition - 4);
if (calculatedChecksum != checksum) {
System.out.println("Checksum " + calculatedChecksum + " didn't match " + checksum);
System.exit(-1);
}
}
}
final BBContainer cont = DBBPool.wrapBB(buf);
DBBPool.registerUnsafeMemory(cont.address());
cont.discard();
}
public void discard() {
buffer.discard();
}
/*
* Prepend length is basically synonymous with writing actual tuple data and not
* the header.
*/
private ListenableFuture<?> write(
final Callable<BBContainer> tupleDataC, final boolean prependLength) {
/*
* Unwrap the data to be written. For the traditional
* snapshot data target this should be a noop.
*/
BBContainer tupleDataTemp;
try {
tupleDataTemp = tupleDataC.call();
/*
* Can be null if the dedupe filter nulled out the buffer
*/
if (tupleDataTemp == null) {
return Futures.immediateFuture(null);
}
} catch (Throwable t) {
return Futures.immediateFailedFuture(t);
}
final BBContainer tupleData = tupleDataTemp;
if (m_writeFailed) {
tupleData.discard();
return null;
}
m_outstandingWriteTasks.incrementAndGet();
Future<BBContainer> compressionTask = null;
if (prependLength) {
BBContainer cont =
DBBPool.allocateDirectAndPool(SnapshotSiteProcessor.m_snapshotBufferCompressedLen);
// Skip 4-bytes so the partition ID is not compressed
// That way if we detect a corruption we know what partition is bad
tupleData.b.position(tupleData.b.position() + 4);
/*
* Leave 12 bytes, it's going to be a 4-byte length prefix, a 4-byte partition id,
* and a 4-byte CRC32C of just the header bytes, in addition to the compressed payload CRC
* that is 16 bytes, but 4 of those are done by CompressionService
*/
cont.b.position(12);
compressionTask = CompressionService.compressAndCRC32cBufferAsync(tupleData.b, cont);
}
final Future<BBContainer> compressionTaskFinal = compressionTask;
ListenableFuture<?> writeTask =
m_es.submit(
new Callable<Object>() {
@Override
public Object call() throws Exception {
try {
if (m_acceptOneWrite) {
m_acceptOneWrite = false;
} else {
if (m_simulateBlockedWrite != null) {
m_simulateBlockedWrite.await();
}
if (m_simulateFullDiskWritingChunk) {
throw new IOException("Disk full");
}
}
int totalWritten = 0;
if (prependLength) {
BBContainer payloadContainer = compressionTaskFinal.get();
try {
final ByteBuffer payloadBuffer = payloadContainer.b;
payloadBuffer.position(0);
ByteBuffer lengthPrefix = ByteBuffer.allocate(12);
m_bytesAllowedBeforeSync.acquire(payloadBuffer.remaining());
// Length prefix does not include 4 header items, just compressd payload
// that follows
lengthPrefix.putInt(payloadBuffer.remaining() - 16); // length prefix
lengthPrefix.putInt(tupleData.b.getInt(0)); // partitionId
/*
* Checksum the header and put it in the payload buffer
*/
PureJavaCrc32C crc = new PureJavaCrc32C();
crc.update(lengthPrefix.array(), 0, 8);
lengthPrefix.putInt((int) crc.getValue());
lengthPrefix.flip();
payloadBuffer.put(lengthPrefix);
payloadBuffer.position(0);
/*
* Write payload to file
*/
while (payloadBuffer.hasRemaining()) {
totalWritten += m_channel.write(payloadBuffer);
}
} finally {
payloadContainer.discard();
}
} else {
while (tupleData.b.hasRemaining()) {
totalWritten += m_channel.write(tupleData.b);
}
}
m_bytesWritten += totalWritten;
m_bytesWrittenSinceLastSync.addAndGet(totalWritten);
} catch (IOException e) {
m_writeException = e;
SNAP_LOG.error(
"Error while attempting to write snapshot data to file " + m_file, e);
m_writeFailed = true;
throw e;
} finally {
try {
tupleData.discard();
} finally {
m_outstandingWriteTasksLock.lock();
try {
if (m_outstandingWriteTasks.decrementAndGet() == 0) {
m_noMoreOutstandingWriteTasksCondition.signalAll();
}
} finally {
m_outstandingWriteTasksLock.unlock();
}
}
}
return null;
}
});
return writeTask;
}
/**
* Process a message pulled off from the network thread, and discard the container once it's
* processed.
*
* @param msg A pair of <sourceHSId, blockContainer>
* @return The restore work, or null if there's no data block to return to the site.
*/
private RestoreWork processMessage(
Pair<Long, Pair<Long, BBContainer>> msg, CachedByteBufferAllocator resultBufferAllocator) {
if (msg == null) {
return null;
}
RestoreWork restoreWork = null;
long hsId = msg.getFirst();
long targetId = msg.getSecond().getFirst();
BBContainer container = msg.getSecond().getSecond();
try {
ByteBuffer block = container.b();
byte typeByte = block.get(StreamSnapshotDataTarget.typeOffset);
final int blockIndex = block.getInt(StreamSnapshotDataTarget.blockIndexOffset);
StreamSnapshotMessageType type = StreamSnapshotMessageType.values()[typeByte];
if (type == StreamSnapshotMessageType.FAILURE) {
VoltDB.crashLocalVoltDB("Rejoin source sent failure message.", false, null);
// for test code only
if (m_expectedEOFs.decrementAndGet() == 0) {
m_EOF = true;
}
} else if (type == StreamSnapshotMessageType.END) {
if (rejoinLog.isTraceEnabled()) {
rejoinLog.trace("Got END message " + blockIndex);
}
// End of stream, no need to ack this buffer
if (m_expectedEOFs.decrementAndGet() == 0) {
m_EOF = true;
}
} else if (type == StreamSnapshotMessageType.SCHEMA) {
rejoinLog.trace("Got SCHEMA message");
block.position(StreamSnapshotDataTarget.contentOffset);
byte[] schemaBytes = new byte[block.remaining()];
block.get(schemaBytes);
m_schemas.put(block.getInt(StreamSnapshotDataTarget.tableIdOffset), schemaBytes);
} else if (type == StreamSnapshotMessageType.HASHINATOR) {
block.position(StreamSnapshotDataTarget.contentOffset);
long version = block.getLong();
byte[] hashinatorConfig = new byte[block.remaining()];
block.get(hashinatorConfig);
restoreWork = new HashinatorRestoreWork(version, hashinatorConfig);
} else {
// It's normal snapshot data afterwards
final int tableId = block.getInt(StreamSnapshotDataTarget.tableIdOffset);
if (!m_schemas.containsKey(tableId)) {
VoltDB.crashLocalVoltDB("No schema for table with ID " + tableId, false, null);
}
// Get the byte buffer ready to be consumed
block.position(StreamSnapshotDataTarget.contentOffset);
ByteBuffer nextChunk = getNextChunk(m_schemas.get(tableId), block, resultBufferAllocator);
m_bytesReceived += nextChunk.remaining();
restoreWork = new TableRestoreWork(tableId, nextChunk);
}
// Queue ack to this block
m_ack.ack(hsId, m_EOF, targetId, blockIndex);
return restoreWork;
} finally {
container.discard();
}
}
void clear() {
BBContainer cont = null;
while ((cont = m_buffers.poll()) != null) {
cont.discard();
}
}
