com.bigdata.ha.pipeline.HAReceiveService Maven / Gradle / Ivy
/**
Copyright (C) SYSTAP, LLC DBA Blazegraph 2006-2016. All rights reserved.
Contact:
SYSTAP, LLC DBA Blazegraph
2501 Calvert ST NW #106
Washington, DC 20008
[email protected]
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program; if not, write to the Free Software
Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
*/
package com.bigdata.ha.pipeline;
import java.io.IOException;
import java.net.BindException;
import java.net.InetSocketAddress;
import java.net.Socket;
import java.nio.ByteBuffer;
import java.nio.channels.AsynchronousCloseException;
import java.nio.channels.SelectionKey;
import java.nio.channels.Selector;
import java.nio.channels.ServerSocketChannel;
import java.nio.channels.SocketChannel;
import java.util.Iterator;
import java.util.Set;
import java.util.concurrent.Callable;
import java.util.concurrent.CancellationException;
import java.util.concurrent.ExecutionException;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.Future;
import java.util.concurrent.FutureTask;
import java.util.concurrent.RejectedExecutionException;
import java.util.concurrent.atomic.AtomicReference;
import java.util.concurrent.locks.Condition;
import java.util.concurrent.locks.Lock;
import java.util.concurrent.locks.ReentrantLock;
import java.util.zip.Adler32;
import org.apache.log4j.Logger;
import com.bigdata.ha.QuorumPipelineImpl;
import com.bigdata.ha.msg.HAMessageWrapper;
import com.bigdata.ha.msg.HASendState;
import com.bigdata.ha.msg.IHAMessage;
import com.bigdata.ha.msg.IHASyncRequest;
import com.bigdata.ha.msg.IHAWriteMessage;
import com.bigdata.ha.msg.IHAWriteMessageBase;
import com.bigdata.ha.pipeline.HASendService.IncSendTask;
import com.bigdata.io.writecache.WriteCache;
import com.bigdata.io.writecache.WriteCacheService;
import com.bigdata.util.BytesUtil;
import com.bigdata.util.ChecksumError;
/**
* Receives data from an {@link HASendService}.
*
* The non-blocking processing of the data cannot proceed until the message
* parameters and an output buffer have been set. So an accept results in a task
* to be run. The Future from this task is returned to the method called from
* the RMI control invocation, thus allowing that method to wait for the
* completion of the data transfer.
*
* @author Martyn Cutcher
* @author Bryan Thompson
*/
public class HAReceiveService extends Thread {
private static final Logger log = Logger
.getLogger(HAReceiveService.class);
/**
* The timeout (milliseconds) on the client {@link Selector}.
* This provides a tradeoff for liveness when responding to
* a pipeline change exception (firstCause) versus spinning
* while awaiting some bytes to read.
*/
static private final long selectorTimeout = 500;
/**
* The timeout (milliseconds) before logging @ WARN that we are
* blocking awaiting data on the socket from the upstream service.
*/
static private final long logTimeout = 10000;
/** The Internet socket address at which this service will listen. */
private final InetSocketAddress addrSelf;
// /**
// * The Internet socket address of a downstream service to which each data
// * transfer will be relayed as it is received (optional and may be
// * null).
// */
// private final InetSocketAddress addrNext;
/**
* Optional callback hook.
*/
private final IHAReceiveCallback callback;
/**
* Used to relay data to a downstream service as it is received. This is
* always allocated, but it will be running iff this service will relay the
* data to a downstream service.
*/
private final HASendService sendService;
public HASendService getSendService() {
return sendService;
}
private final ExecutorService executor = Executors
.newSingleThreadExecutor();
// private ServerSocketChannel server;
// private FutureTask readFuture;
/**
* Service run state enumeration.
* @author Bryan Thompson
*/
private static enum RunState {
Start(0), Running(1), ShuttingDown(2), Shutdown(3);
private RunState(final int level) {
this.level = level;
}
@SuppressWarnings("unused")
private final int level;
}
/*
* The lock and the things which it guards.
*/
/**
* The {@link Lock}.
*/
private final Lock lock = new ReentrantLock();
/**
* {@link Condition} signaled when the {@link #waitFuture} is ready.
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)} awaits this
* {@link Condition}. Once signaled, it returns the {@link #waitFuture} to
* the caller and clears {@link #waitFuture} to null.
*
* The {@link Condition}s {@link #messageReady} and {@link #futureRead}
* respectively manage the hand off of the message (to the {@link ReadTask})
* and the {@link #waitFuture} (to the thread calling
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)}.
*/
private final Condition futureReady = lock.newCondition();
/**
* {@link Condition} signaled when a new {@link IHAWriteMessage} has been
* set on {@link #message} by
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)}.
*/
private final Condition messageReady = lock.newCondition();
/**
* {@link RunState} for the {@link HAReceiveService}. This is used to manage
* startup and termination state transitions.
*/
private RunState runState = RunState.Start;
/**
* The current {@link IHAWriteMessageBase}. This message provides metadata
* about the expected buffer transfer. This field is set by
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)}.
*/
private M message;
/**
* The current receive buffer. This buffer is populated with data based on
* the expected {@link IHAWriteMessage#getSize()}. The data is verified by
* comparing the checksum of the buffer to the expected checksum as
* specified by {@link IHAWriteMessage#getChk()}.
*/
private ByteBuffer localBuffer;
/**
* {@link Future} for the current buffer transfer used to await the
* termination of that transfer by the {@link ReadTask}.
*
* Note: The {@link #readFuture} is cleared to null as soon as
* the buffer transfer is complete.
*/
private FutureTask readFuture;
/**
* {@link Future} for the current buffer transfer used to await the
* termination of that transfer by the thread that calls
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)}.
*
* Note: The {@link #waitFuture} is cleared to null as soon as
* it is returned to the caller. This can occur before the buffer transfer
* is complete. Therefore, {@link ReadTask} MUST NOT wait on the
* {@link #waitFuture}.
*/
private FutureTask waitFuture;
/**
* The Internet socket address of a downstream service to which each data
* transfer will be relayed as it is received (optional and may be
* null).
*
* Note: This an {@link AtomicReference} for visibility in
* {@link #toString()}, which does not obtain the {@link #lock}. The
* {@link AtomicReference} also make changes in the downstream service
* address visible inside of {@link ReadTask}.
*/
private final AtomicReference addrNextRef;
/**
* Private buffer used to incrementally compute the checksum of the data as
* it is received. The purpose of this buffer is to take advantage of more
* efficient bulk copy operations from the NIO buffer into a local byte[] on
* the Java heap against which we then track the evolving checksum of the
* data.
*/
private final byte[] heapBuffer = new byte[512];
/*
* Note: toString() implementation is non-blocking.
*/
@Override
public String toString() {
return super.toString() + "{addrSelf=" + addrSelf + ", addrNext="
+ addrNextRef.get() + "}";
}
/** The Internet socket address at which this service will listen (immutable) */
public InetSocketAddress getAddrSelf() {
return addrSelf;
}
/**
* The Internet socket address to which this service will relay messages
* (dynamic and MAY be null).
*
* @see #changeDownStream(InetSocketAddress)
*/
public InetSocketAddress getAddrNext() {
return addrNextRef.get();
}
/**
* Create a new service instance - you MUST {@link Thread#start()} the
* service.
*
* Note: addrNext can be changed dynamically.
*
* @param addrSelf
* The Internet socket address at which this service will listen.
* @param addrNext
* The Internet socket address of a downstream service to which
* each data transfer will be relayed as it is received
* (optional).
*
* @see #changeDownStream(InetSocketAddress)
*/
public HAReceiveService(final InetSocketAddress addrSelf,
final InetSocketAddress addrNext) {
this(addrSelf, addrNext, null/* callback */);
}
/**
* Create a new service instance - you MUST {@link Thread#start()} the
* service.
*
* @param addrSelf
* The Internet socket address at which this service will listen.
* @param addrNext
* The Internet socket address of a downstream service to which
* each data transfer will be relayed as it is received
* (optional).
* @param callback
* An object which will be notified as each payload arrives.
*/
public HAReceiveService(final InetSocketAddress addrSelf,
final InetSocketAddress addrNext,
final IHAReceiveCallback callback) {
if (addrSelf == null)
throw new IllegalArgumentException();
this.addrSelf = addrSelf;
this.addrNextRef = new AtomicReference(addrNext);
this.callback = callback;
// Note: Always allocate since the addrNext can change.
this.sendService = new HASendService();
// Thread will not prevent JVM exit.
setDaemon(true);
// Give the thread a useful name.
setName(HAReceiveService.class.getName() + "@" + hashCode()
+ "{addrSelf=" + addrSelf + "}");
if (log.isInfoEnabled())
log.info("Created: " + this);
}
/**
* Extended to {@link #terminate()} processing in order to ensure that
* the service is eventually shutdown.
*/
@Override
protected void finalize() throws Throwable {
terminate();
super.finalize();
}
/**
* Immediate shutdown.
*/
public void terminate() {
lock.lock();
try {
switch (runState) {
case ShuttingDown: // already shutting down.
case Shutdown: // already shutdown.
return;
default:
runState = RunState.ShuttingDown;
this.interrupt();
}
} finally {
lock.unlock();
}
if (sendService != null)
sendService.terminate();
executor.shutdownNow();
}
/**
* Block until the service is shutdown.
*/
public void awaitShutdown() throws InterruptedException {
/*
* Wait until we observe that the service is no longer running while
* holding the lock.
*
* Note: When run() exits it MUST signalAll() on both [futureReady] and
* [messageReady] so that all threads watching those Conditions notice
* that the service is no longer running.
*/
lock.lockInterruptibly();
try {
while (true) {
switch (runState) {
case Start:
case Running:
case ShuttingDown:
futureReady.await();
continue;
case Shutdown:
// Exit terminate().
return;
default:
throw new AssertionError();
}
}
} finally {
lock.unlock();
}
}
public void start() {
super.start();
lock.lock();
try {
// Wait for state change from Start
while (runState == RunState.Start) {
try {
futureReady.await();
} catch (InterruptedException e) {
// let's go around again
}
}
} finally {
lock.unlock();
}
}
public void run() {
lock.lock();
try {
// Change the run state and signal anyone who might be watching.
runState = RunState.Running;
futureReady.signalAll();
messageReady.signalAll();
} finally {
lock.unlock();
}
ServerSocketChannel server = null;
try {
/*
* Open a non-blocking server socket channel and start listening.
*/
server = ServerSocketChannel.open();
{
/*
* Robustly attempt to bind the address and port where this
* service will listen.
*
* Note: The retry is here because the port is not freed up
* immediately when we close the existing socket connection
*/
boolean didBind = false;
for (int i = 0; i < 3; i++) {
try {
server.socket().bind(addrSelf);
didBind = true;
break;
} catch (BindException ex) {
log.warn("Sleeping to retry: " + ex);
Thread.sleep(100/* ms */);
continue;
}
}
if (!didBind) {
server.socket().bind(addrSelf);
}
}
server.configureBlocking(false);
if(log.isInfoEnabled())
log.info("Listening on: " + addrSelf);
runNoBlock(server);
} catch (InterruptedException e) {
/*
* @todo what is the normal shutdown exception?
*/
log.info("Shutdown");
} catch (Throwable t) {
log.error(t, t);
throw new RuntimeException(t);
} finally {
if (server != null) {
try {
server.close();
} catch (IOException e) {
log.error(e, e);
}
}
lock.lock();
try {
runState = RunState.Shutdown;
messageReady.signalAll();
futureReady.signalAll();
} finally {
lock.unlock();
}
}
}
/**
* The client socket connection that was obtained when we accepted the
* upstream reader.
*
* Note: The {@link Client} connection is reused across {@link ReadTask}s.
*
* Note: Exposed to {@link #changeUpStream()}.
*/
private final AtomicReference clientRef = new AtomicReference(null);
/**
* Loops accepting requests and scheduling {@link ReadTask}s. Note that a
* local caller must hand us a buffer and {@link IHAWriteMessageBase} using
* {@link #receiveData(IHAWriteMessageBase, ByteBuffer)} before we will
* accept data on the {@link SocketChannel}.
*
* @throws IOException
* @throws ExecutionException
* @throws InterruptedException
*/
private void runNoBlock(final ServerSocketChannel server) throws IOException,
InterruptedException, ExecutionException {
try {
while (true) {
// wait for the message to be set (actually, msg + buffer).
lock.lockInterruptibly();
try {
// wait for the message.
while (message == null) {
switch (runState) {
case Running:
break;
case ShuttingDown:
// Service is terminating.
return;
case Start:
case Shutdown:
default:
throw new AssertionError(runState.toString());
}
messageReady.await();
}
// Note the message.
final M msg = message;
// Message cleared.
message = null;
// Setup task to read buffer for that message.
readFuture = waitFuture = new FutureTask(
new ReadTask(server, clientRef, msg,
localBuffer, heapBuffer, sendService,
addrNextRef, callback));
// [waitFuture] is available for receiveData().
futureReady.signalAll();
} finally {
lock.unlock();
}
/*
* The ReadTask now listens for the accept, ensuring that a
* future is available as soon as a message is present.
*/
try {
executor.execute(readFuture);
} catch (RejectedExecutionException ex) {
readFuture.cancel(true/* mayInterruptIfRunning */);
log.error(ex);
}
/*
* Note: We might have to wait for the Future to avoid having
* more than one ReadTask at a time, but we should log and
* ignore any exception and restart the loop.
*
* The loop needs to keep running. The thread that called
* receiveData() will return the [waitFuture] and will notice
* any exception through that Future.
*/
try {
readFuture.get();
} catch (Exception e) {
log.error(e, e);
}
lock.lockInterruptibly();
try {
readFuture = null;
} finally {
lock.unlock();
}
} // while(true)
} finally {
final Client client = clientRef.get();
if (client != null) {
client.close();
}
}
}
/**
* Class encapsulates the connection state for the socket channel used to
* receive from on the upstream {@link HASendService}.
*
* @author Bryan
* Thompson
*/
static private class Client {
private final SocketChannel client;
private final Selector clientSelector;
private final SelectionKey clientKey;
/**
* When a pipeline change event is handled, we need to throw out an
* exception rather than just cancelling the
* {@link HAReceiveService#readFuture}. Cancelling the
* {@link HAReceiveService#readFuture} causes a
* {@link CancellationException} to be propoagated back to the remote
* service that invoked
* {@link IPipelineGlue#receiveAndReplicate(IHASyncRequest, IHAWriteMessage)}
* . That {@link CancellationException} gets interpreted as a normal
* termination in {@link QuorumPipelineImpl} and results in the
* retrySend() logic NOT retrying and resending and thus breaks the
* robustness of write pipeline replication.
*
* Instead, the pipeline change events are used to set a
* {@link Throwable} that is then thrown out of
* {@link ReadTask#doReceiveAndReplicate(Client)} and thus appears as a
* non-normal termination of the read future in the upstream service.
* This allows retrySend() to do the right thing - namely it sends an
* RMI message to the new downstream service and retransmits the payload
* along the write pipeline.
*/
private final AtomicReference firstCause = new AtomicReference();
// /** Used to replicate the message to the downstream service (if any). */
// private final HASendService downstream;
/**
* Gets the client connection and open the channel in a non-blocking
* mode so we will read whatever is available and loop until all data
* has been read.
*/
public Client(//
final ServerSocketChannel server //
// , final HASendService downstream //
// , final InetSocketAddress addrNext//
) throws IOException {
try {
/*
* Note: This binds a port for a specific upstream HASendService
* that will be talking to this HAReceiveService.
*/
client = server.accept();
client.configureBlocking(false);
if (!client.finishConnect())
throw new IOException("Upstream client not connected");
clientSelector = Selector.open();
// must register OP_READ selector on the new client
clientKey = client.register(clientSelector,
SelectionKey.OP_READ);
if (log.isInfoEnabled())
log.info("Accepted new connection");
// this.downstream = downstream;
//
// // Prepare downstream (if any) for incremental transfers
// if (addrNext != null) {
//
// downstream.start(addrNext);
//
// }
} catch (IOException ex) {
close();
throw ex;
}
}
@Override
public String toString() {
final Socket s = client.socket();
return super.toString() //
+ "{client.isOpen()=" + client.isOpen()//
+ ",client.isConnected()=" + client.isConnected()//
+ ",socket.isInputShutdown()="
+ (s == null ? "N/A" : s.isInputShutdown())//
+ ",clientSelector.isOpen=" + clientSelector.isOpen()//
+ "}";
}
private void close() throws IOException {
if (log.isInfoEnabled())
log.info("Closing client connection: " + this);
clientKey.cancel();
try {
client.close();
} finally {
// try {
clientSelector.close();
// } finally {
// if (downstream != null) {
// downstream.terminate();
// }
// }
}
}
/**
* Wraps {@link SocketChannel#read(ByteBuffer)} to test for an EOF and
* calls {@link #close()} if an EOF is reached.
*
* @param dst
* The destination buffer.
*
* @return The #of bytes read.
*
* @throws IOException
*/
private int read(final ByteBuffer dst) throws IOException {
final int rdlen = client.read(dst);
if (rdlen == -1) {
close();
}
return rdlen;
}
/**
* Termination path used to signal a pipeline change through exception
* control back to the leader. The leader will then handle this in
* {@link QuorumPipelineImpl}'s retrySend() method.
*/
private void checkFirstCause() throws RuntimeException {
final Throwable t = firstCause.getAndSet(null);
if (t != null) {
try {
close();
} catch (IOException ex) {
log.warn(ex, ex);
}
throw new RuntimeException(t);
}
}
}
/**
* Read task is called with a {@link ServerSocketChannel}, a message
* describing the data to be received, and a buffer into which the data will
* be copied. It waits for the client connection and then copies the data
* into the buffer, computing the checksum as it does, and optionally
* transfer the data onto the downstream {@link SocketChannel}.
*
* @author Bryan
* Thompson
*
* @todo report counters
*
* report the #of chunks per payload so we can decide if the private
* byte[] for computing the checksum is a good size.
*
* report the #of payloads.
*/
static private class ReadTask implements
Callable {
private final ServerSocketChannel server;
private final AtomicReference clientRef;
private final M message;
private final ByteBuffer localBuffer;
/**
* Used to transfer received data to the downstream service (if any).
*/
private final HASendService sendService;
/**
* The address of the downstream service -or- null iff
* there is no downstream service.
*/
private final AtomicReference addrNextRef;
/**
* Optional callback.
*/
private final IHAReceiveCallback callback;
private final Adler32 chk = new Adler32();
private final byte[] heapBuffer;
/**
*
* @param server
* @param clientRef
* The client socket, selector, etc.
* @param message
* The message carrying metadata about the data to be
* received (especially its byte length and its
* {@link Adler32} checksum).
* @param localBuffer
* The buffer into which the data will be transferred.
* @param downstream
* The {@link HASendService} used to relay data to the
* downstream node.
* @param addrNext
* An {@link AtomicReference} for address of the downstream
* node. The value within that {@link AtomicReference} may be
* updated by
* {@link HAReceiveService#changeDownStream(InetSocketAddress)}
* . That value will be null if this is the last
* node in the write pipeline at the time the value is
* observed.
* @param callback
* An optional callback.
*/
public ReadTask(final ServerSocketChannel server,
final AtomicReference clientRef, final M message,
final ByteBuffer localBuffer, final byte[] heapBuffer,
final HASendService downstream,
final AtomicReference addrNextRef,
final IHAReceiveCallback callback) {
if (server == null)
throw new IllegalArgumentException();
if (clientRef == null)
throw new IllegalArgumentException();
if (message == null)
throw new IllegalArgumentException();
if (heapBuffer == null)
throw new IllegalArgumentException();
if (localBuffer == null)
throw new IllegalArgumentException();
if (downstream == null)
throw new IllegalArgumentException();
this.server = server;
this.clientRef = clientRef;
this.message = message;
this.localBuffer = localBuffer;
this.heapBuffer = heapBuffer;
this.sendService = downstream;
this.addrNextRef = addrNextRef;
this.callback = callback;
}
/**
* Blocking wait for a client connection.
*
* @throws IOException
* if something goes wrong.
*/
protected void awaitAccept() throws IOException {
// blocking wait for a client connection.
final Selector serverSelector = Selector.open();
try {
final SelectionKey serverKey = server.register(serverSelector,
SelectionKey.OP_ACCEPT);
try {
serverSelector.select(); // blocks
final Set keys = serverSelector
.selectedKeys();
final Iterator iter = keys.iterator();
while (iter.hasNext()) {
final SelectionKey key = (SelectionKey) iter.next();
iter.remove();
if (key != serverKey)
throw new AssertionError();
break;
}
} finally {
serverKey.cancel();
}
} finally {
serverSelector.close();
}
}
/**
* Update the running checksum. This uses the {@link #heapBuffer} to
* amoritize the cost of the transfers for the incremental checksum
* maintenance.
*
* @param rdlen
* The #of bytes read in the last read from the socket into
* the {@link #localBuffer}.
*/
private void updateChk(final int rdlen) {
// isolate changes to (pos,limit).
final ByteBuffer b = localBuffer.asReadOnlyBuffer();
// current position (and limit of how much data we need to chksum).
final int mark = b.position();
// rewind to the first byte to be read.
b.position(mark - rdlen);
for (int pos = mark - rdlen; pos < mark; pos += heapBuffer.length) {
// #of bytes to copy into the local byte[].
final int len = Math.min(mark - pos, heapBuffer.length);
// copy into Java heap byte[], advancing b.position().
b.get(heapBuffer, 0/* off */, len);
// update the running checksum.
chk.update(heapBuffer, 0/* off */, len);
}
}
@Override
public Void call() throws Exception {
try {
return doInnerCall();
} catch (Throwable t) {
/*
* Log anything thrown out of this task. We check the Future of
* this task, but that does not tell us what exception is thrown
* in the Thread executing the task when the Future is cancelled
* and that thread is interrupted. In particular, we are looking
* for the InterruptedException, ClosedByInterruptException,
* etc.
*/
log.error(
"client="
+ clientRef.get()
+ ", msg="
+ message
+ ", marker="
+ HASendState.decode(message.getHASendState()
.getMarker()) + ", cause=" + t, t);
if (t instanceof Exception)
throw (Exception) t;
if (t instanceof RuntimeException)
throw (RuntimeException) t;
throw new RuntimeException(t);
}
}
private Void doInnerCall() throws Exception {
// awaitAccept();
//
// /*
// * Get the client connection and open the channel in a non-blocking
// * mode so we will read whatever is available and loop until all
// * data has been read.
// */
// final SocketChannel client = server.accept();
// client.configureBlocking(false);
//
// final Selector clientSelector = Selector.open();
//
// // must register OP_READ selector on the new client
// final SelectionKey clientKey = client.register(clientSelector,
// SelectionKey.OP_READ);
Client client = clientRef.get();
// if (client != null) {
//
// /*
// * Note: We need to know when the client connection is no longer
// * valid. The code here does not appear to do the trick.
// * changeUpStream() is handling this instead.
// *
// * We need to decide whether the client is no longer valid
// * (either because the upstream HASendService has changed (our
// * predecessor in the pipeline might have died) or because it
// * has closed is socket connection to this HAReceiveService).
// *
// * Either way, we need to establish a client connection using
// * awaitAccept().
// */
// if (!client.client.isConnected()) {
// log.warn("Closing old client connection.");
// clientRef.set(client = null);
// }
//
// }
if (client == null || !client.client.isOpen()
|| !client.clientSelector.isOpen()) {
log.warn("Re-opening upstream client connection");
final Client tmp = clientRef.getAndSet(null);
if (tmp != null) {
// Close existing connection if not open.
tmp.close();
}
/*
* Accept and the initialize a connection from the upstream
* HASendService.
*/
// Accept a client connection (blocks)
awaitAccept();
// New client connection.
client = new Client(server);//, sendService, addrNext);
// save off reference.
clientRef.set(client);
}
// boolean success = false;
// try {
doReceiveAndReplicate(client);
// success = true;
// success.
return null;
// } finally {
// try {
// if(success) {
// ack(client);
// } else {
// nack(client);
// }
// } catch (IOException ex) {
// // log and ignore.
// log.error(ex, ex);
// }
// }
} // call.
private void doReceiveAndReplicate(final Client client)
throws Exception {
// /**
// * The first cause if downstream replication fails. We make a note
// * of this first cause, continue to drain the payload, and then
// * rethrow the first cause once the payload has been fully drained.
// * This is necessary to ensure that the socket channel does not have
// * partial data remaining from an undrained payload.
// *
// * @see HA wire pulling and sure kill testing
// *
// * Note: It appears that attempting to drain the
// * payload is risky since there are a variety of ways in which
// * the process might be terminated. It seems to be safer to
// * drain the socket channel until we reach a marker that gives
// * us confidence that we are at the payload for the message
// * that is being processed.
// */
// Throwable downstreamFirstCause = null;
/*
* We should now have parameters ready in the WriteMessage and can
* begin transferring data from the stream to the writeCache.
*/
final long begin = System.currentTimeMillis();
long mark = begin;
// #of bytes remaining (to be received).
int rem = message.getSize();
// End of stream flag.
boolean EOS = false;
// for debug retain number of low level reads
int reads = 0;
final DrainToMarkerUtil drainUtil = message.getHASendState() != null ? new DrainToMarkerUtil(
message.getHASendState().getMarker(), client) : null;
while (rem > 0 && !EOS) {
// block up to the timeout.
final int nkeys = client.clientSelector
.select(selectorTimeout/* ms */);
// Check for termination (first cause exception).
client.checkFirstCause();
if (nkeys == 0) {
/*
* Nothing available.
*/
// time since last mark.
final long now = System.currentTimeMillis();
final long elapsed = now - mark;
if (elapsed > logTimeout) {
// Issue warning if we have been blocked for a while.
log.warn("Blocked: awaiting " + rem + " out of "
+ message.getSize() + " bytes.");
mark = now;// reset mark.
}
if (!client.client.isOpen()
|| !client.clientSelector.isOpen()) {
/*
* The channel has been closed. The request must be
* failed. TODO Or set EOF:=true?
*
* Note: The [callback] is NOT notified. The service
* that issued the RMI request to this service to
* receive the payload over the HAReceivedService will
* see this exception thrown back across the RMI
* request.
*
* @see HAReceiveService.receiveData().
*/
throw new AsynchronousCloseException();
}
// no keys. nothing to read.
continue;
}
final Set keys = client.clientSelector
.selectedKeys();
final Iterator iter = keys.iterator();
while (iter.hasNext()) {
// Check for termination.
client.checkFirstCause();
iter.next();
iter.remove();
if (!drainUtil.findMarker()) {
// continue to drain until the marker.
continue;
}
final int rdlen = client.read(localBuffer);
if (log.isTraceEnabled())
log.trace("Read " + rdlen + " bytes with "
+ (rdlen > 0 ? rem - rdlen : rem)
+ " bytes remaining.");
if (rdlen > 0) {
reads++;
updateChk(rdlen);
}
if (rdlen == -1) {
// The stream is closed?
EOS = true;
break;
}
rem -= rdlen;
if (callback != null) {
// notify of incremental read.
callback.incReceive(message, reads, rdlen, rem);
}
forwardReceivedBytes(client, rdlen);
} // while(itr.hasNext())
} // while( rem > 0 && !EOS )
if (localBuffer.position() != message.getSize())
throw new IOException("Receive length error: rem=" + rem
+ ", EOS=" + EOS + ", localBuffer.pos="
+ localBuffer.position() + ", message.size="
+ message.getSize());
// prepare for reading.
localBuffer.flip();
if (log.isTraceEnabled())
log.trace("Prior check checksum: " + chk.getValue()
+ " for position: " + localBuffer.position()
+ ", limit: " + localBuffer.limit()
+ ", number of reads: " + reads + ", buffer: "
+ localBuffer);
if (message.getChk() != (int) chk.getValue()) {
throw new ChecksumError("msg=" + message.toString()
+ ", actual=" + (int) chk.getValue());
}
// Check for termination.
client.checkFirstCause();
if (callback != null) {
/*
* The message was received and (if there is a downstream
* service) successfully replicated to the downstream service.
* We now invoke the callback to given this service an
* opportunity to handle the message and the fully received
* payload.
*/
callback.callback(message, localBuffer);
}
} // call()
/**
* Forward the most recent transfer bytes downstream.
*
*
* Note: [addrNext] is final. If the downstream address is changed, then
* the {@link ReadTask} is interrupted using its {@link Future} and the
* WriteCacheService on the leader will handle the error by
* retransmitting the current cache block.
*
* The rdlen is checked for non zero to avoid an
* IllegalArgumentException.
*
* Note: loop since addrNext might change asynchronously.
*
* @throws ExecutionException
* @throws InterruptedException
* @throws ImmediateDownstreamReplicationException
*
* @todo Since the downstream writes are against a blocking mode
* channel, the receiver on this node runs in sync with the
* receiver on the downstream node. In fact, those processes could
* be decoupled with a bit more effort and are only required to
* synchronize by the end of each received payload.
*
* @see
* HA wire pulling and sure kill testing
*/
private void forwardReceivedBytes(final Client client, final int rdlen)
throws InterruptedException, ExecutionException,
ImmediateDownstreamReplicationException {
while (true) {
if (rdlen != 0 && addrNextRef.get() != null) {
if (log.isTraceEnabled())
log.trace("Incremental send of " + rdlen + " bytes");
final ByteBuffer out = localBuffer.asReadOnlyBuffer();
out.position(localBuffer.position() - rdlen);
out.limit(localBuffer.position());
synchronized (sendService) {
/*
* Note: Code block is synchronized on [downstream] to
* make the decision to start the HASendService that
* relays to [addrNext] atomic. The HASendService uses
* [synchronized] for its public methods so we can
* coordinate this lock with its synchronization API.
*/
if (!sendService.isRunning()) {
/*
* Prepare send service for incremental transfers to
* the specified address.
*/
// Check for termination.
client.checkFirstCause();
// Note: use then current addrNext!
sendService.start(addrNextRef.get());
continue;
}
}
// Check for termination.
client.checkFirstCause();
/*
* Send and await Future. If this is the first chunk of a
* payload and a marker exists, then send the marker as
* well.
*/
sendService
.send(out,
out.position() == 0
&& message.getHASendState() != null ? message
.getHASendState().getMarker()
: null).get();
}
break; // break out of the inner while loop.
} // while(true)
}
} // class ReadTask
/**
* Helper class to drain bytes from the upstream socket until we encounter a
* marker in the stream that immediately proceeds the desired payload.
*
* @author Martyn
* Cutcher
*
* @see HA
* wire pulling and sure kill testing
*/
static private class DrainToMarkerUtil {
final private byte[] marker;
final private byte[] markerBuffer;
final private ByteBuffer markerBB;
final private Client client;
private boolean foundMarkerInInitialPosition = true;
private int markerIndex = 0;
private int nreads = 0;
private int nmarkerbytematches = 0;
private long bytesRead = 0L;
DrainToMarkerUtil(final byte[] marker, final Client client) {
this.marker = marker;
this.markerBuffer = marker == null ? null : new byte[marker.length];
this.markerBB = marker == null ? null : ByteBuffer
.wrap(markerBuffer);
this.client = client;
if (log.isDebugEnabled())
log.debug("Receive token: " + BytesUtil.toHexString(marker));
}
/**
* Note that the logic for finding the token bytes depends on the first
* byte in the token being unique!
*
* We have to be a bit clever to be sure we do not read beyond the token
* and therefore complicate the reading into the localBuffer.
*
* This is optimized for the normal case where the marker is read as
* from the next bytes from the stream. In the worst case scenario this
* could read large amounts of data only a few bytes at a time, however
* this is not in reality a significant overhead.
*/
boolean findMarker() throws IOException {
if (markerIndex == marker.length) {
// Marker already found for this payload.
return true;
}
if (log.isDebugEnabled())
log.debug("Looking for token, " + BytesUtil.toHexString(marker)
+ ", reads: " + nreads);
while (markerIndex < marker.length) {
final int remtok = marker.length - markerIndex;
markerBB.limit(remtok);
markerBB.position(0);
final int rdLen = client.read(markerBB);
if (rdLen == -1) {
throw new IOException("EOF: nreads=" + nreads
+ ", bytesRead=" + bytesRead);
}
nreads++;
bytesRead += rdLen;
for (int i = 0; i < rdLen; i++) {
if (markerBuffer[i] != marker[markerIndex]) {
if (foundMarkerInInitialPosition) {
/*
* The marker was not found in the initial position
* in the stream. We are going to drain data until
* we can match the marker.
*/
foundMarkerInInitialPosition = false;
log.error("Marker not found: skipping");
}
markerIndex = 0;
if (markerBuffer[i] == marker[markerIndex]) {
markerIndex++;
}
} else {
markerIndex++;
nmarkerbytematches++;
}
}
if (nreads % 10000 == 0) {
if (log.isDebugEnabled())
log.debug("...still looking: reads=" + nreads
+ ", bytesRead=" + bytesRead);
}
}
if (markerIndex != marker.length) {
/*
* Partial marker has been read, but we do not have enough data
* for a full match yet.
*/
if (log.isDebugEnabled())
log.debug("Not found token yet!");
return false;
} else {
if (log.isDebugEnabled())
log.debug("Found token after " + nreads
+ " token reads and " + nmarkerbytematches
+ " byte matches");
return true;
}
}
}
/**
* Receive data into the caller's buffer as described by the caller's
* message.
*
* @param msg
* The metadata about the data to be transferred.
* @param buffer
* The buffer in which this service will receive the data. The
* buffer MUST be large enough for the data to be received. The
* buffer SHOULD be a direct {@link ByteBuffer} in order to
* benefit from NIO efficiencies. This method will own the buffer
* until the returned {@link Future} is done.
*
* @return A {@link Future} which you can await. The {@link Future} will
* become available when the data has been transferred into the
* buffer, at which point the position will be ZERO (0) and the
* limit will be the #of bytes received into the buffer. If the data
* transfer fails or is interrupted, the {@link Future} will report
* the exception.
*
* @throws InterruptedException
*/
public Future receiveData(final M msg, final ByteBuffer buffer)
throws InterruptedException {
if (msg == null)
throw new IllegalArgumentException();
if (buffer == null)
throw new IllegalArgumentException();
lock.lockInterruptibly();
try {
assert message == null;
message = msg;
localBuffer = buffer;// DO NOT duplicate()! (side-effects required)
localBuffer.limit(message.getSize());
localBuffer.position(0);
messageReady.signalAll();
if (log.isTraceEnabled())
log.trace("Will accept data for message: msg=" + msg);
while (waitFuture == null) {
switch (runState) {
case Start:
case Running:
// fall through and await signal.
break;
case ShuttingDown:
case Shutdown:
throw new RuntimeException("Service closed.");
default:
throw new AssertionError();
}
// await signal.
futureReady.await();
}
assert waitFuture != null;
return waitFuture;
// return readFuture; // Note: readFuture observed as null (!)
} finally {
waitFuture = null;
lock.unlock();
}
}
/**
* Hook to notice receive events.
*
* @author Bryan Thompson
*
* @param
*/
public interface IHAReceiveCallback {
/**
* Notify that some payload bytes have been incrementally received for
* an {@link IHAMessage}. This is invoked each time some data has been
* read from the upstream socket.
*
* @param msg
* The message.
* @param nreads
* The number of reads performed against the upstream socket
* for this message.
* @param rdlen
* The number of bytes read from the socket in this read.
* @param rem
* The number of bytes remaining before the payload has been
* fully read.
*
* @throws Exception
*/
void incReceive(M msg, int nreads, int rdlen, int rem) throws Exception;
/**
* Hook invoked once a buffer has been received.
*
* @param msg
* The message.
* @param data
* The buffer containing the data. The position() will be
* ZERO (0). The limit() will be the #of bytes available. The
* implementation MAY have side effects on the buffer state
* (position, limit, etc).
*
* @throws Exception
*/
void callback(M msg, ByteBuffer data) throws Exception;
}
/**
* Change the address to which the payloads are being relayed. This
* terminates the embedded {@link HASendService}. The {@link HASendService}
* will be restarted with the new {@link InetSocketAddress} (if any) by the
* {@link ReadTask}.
*
* Note: The {@link ReadTask} will throw out an exception when if there was
* a downstream target when the {@link IncSendTask} is interrupted. Since
* the {@link ReadTask} lacks the context to issue the appropriate RMI to
* the downstream task, the exception must be caught and handled by the
* {@link WriteCacheService}. It can simply rediscover the new downstream
* service and then re-submit both the RMI and the {@link WriteCache} block.
*
* @param addrNext
* The new address -or- null if payloads should not
* be relayed at this time.
*/
public void changeDownStream(final InetSocketAddress addrNext) {
lock.lock();
try {
if (log.isInfoEnabled())
log.info("addrNext(old)=" + this.addrNextRef.get()
+ ", addrNext(new)=" + addrNext + ", readFuture="
+ readFuture);
final Client c = clientRef.get();
if (c != null && readFuture != null) {
/*
* Set firstCause. doReceiveAndReplicate() will notice this and
* throw the (wrapped) exception back to the caller. This allows
* retrySend() on the leader to differentiate between normal
* termination of a downstream service and a pipeline change
* event.
*
* Note: We do this *instead* of interrupting the [readFuture].
* The cause will be thrown out after a timeout on the client
* Selector or the next time any bytes are received at that
* Selector.
*
* Note: The code path that interrupted the [readFuture] would
* only do so if the [readFuture] was non-null. The same
* behavior is preserved here. This subtlty means that a
* pipeline change event that occurs *before* the next attempt
* to receive a payload will succeed while a change that occurs
* once we have started to read data will fail.
*/
c.firstCause.set(new PipelineDownstreamChange());
}
// if (readFuture != null) {
//
// // Interrupt the current receive operation.
// readFuture.cancel(true/* mayInterruptIfRunning */);
//
// }
synchronized (sendService) {
if (sendService.isRunning()) {
// Terminate HASendService (iff running).
sendService.terminate();
}
/*
* Save the new addr.
*
* Note: We need to do this while holding the monitor for the
* [sendService] since the update must be visible if we restart
* the sendService.
*/
this.addrNextRef.set(addrNext);
}
/*
* Note: Do not start the service here. It will be started by the
* next ReadTask, which will have the then current value of addrNext.
*/
// if (addrNext != null) {
//
// // Start send service w/ a new connection.
// downstream.start(addrNext);
//
// }
} finally {
lock.unlock();
}
}
/**
* Method must be invoked when the upstream service is changed. The method
* is responsible for interrupting the current {@link RunTask} (if any) and
* closing the client socket connection that was used to receive data from
* the upstream service. A new connection will be accepted by the next
* {@link RunTask}.
*/
public void changeUpStream() {
lock.lock();
try {
if (log.isInfoEnabled())
log.info("");
final Client oldClient = clientRef.getAndSet(null);
if (oldClient != null) {
log.warn("Cleared Client reference.");
}
if (oldClient != null && readFuture != null) {
/*
* Set firstCause. doReceiveAndReplicate() will notice this and
* throw the (wrapped) exception back to the caller. This allows
* retrySend() on the leader to differentiate between normal
* termination of a downstream service and a pipeline change
* event.
*
* Note: We do this *instead* of interrupting the [readFuture].
* The cause will be thrown out after a timeout on the client
* Selector or the next time any bytes are received at that
* Selector.
*
* Note: The code path that interrupted the [readFuture] would
* only do so if the [readFuture] was non-null. The same
* behavior is preserved here. This subtlty means that a
* pipeline change event that occurs *before* the next attempt
* to receive a payload will succeed while a change that occurs
* once we have started to read data will fail.
*/
oldClient.firstCause.set(new PipelineUpstreamChange());
}
// if (readFuture != null) {
//
// // Interrupt the current receive operation.
// readFuture.cancel(true/* mayInterruptIfRunning */);
//
// }
/*
* Explicitly close the client socket channel.
*/
{
// final Client oldClient = clientRef.getAndSet(null);
if (oldClient != null) {
if (log.isInfoEnabled())
log.info("Closing client connection");
try {
oldClient.client.close();
} catch (IOException e) {
log.warn(e, e);
}
}
}
} finally {
lock.unlock();
}
}
}