/**
* Starts communication.
*
* @throws IgniteCheckedException If failed.
*/
public void start() throws IgniteCheckedException {
try {
locHost = U.getLocalHost();
} catch (IOException e) {
throw new IgniteCheckedException("Failed to initialize local address.", e);
}
try {
shmemSrv = resetShmemServer();
} catch (IgniteCheckedException e) {
U.warn(log, "Failed to start shared memory communication server.", e);
}
try {
// This method potentially resets local port to the value
// local node was bound to.
nioSrvr = resetNioServer();
} catch (IgniteCheckedException e) {
throw new IgniteCheckedException("Failed to initialize TCP server: " + locHost, e);
}
locProcDesc.address(locHost.getHostAddress());
locProcDesc.sharedMemoryPort(boundTcpShmemPort);
locProcDesc.tcpPort(boundTcpPort);
locIdMsg = new ProcessHandshakeMessage(locProcDesc);
if (shmemSrv != null) {
shmemAcceptWorker = new ShmemAcceptWorker(shmemSrv);
new IgniteThread(shmemAcceptWorker).start();
}
nioSrvr.start();
}
/**
* Sends message to Hadoop process.
*
* @param desc
* @param msg
* @throws IgniteCheckedException
*/
public void sendMessage(HadoopProcessDescriptor desc, HadoopMessage msg)
throws IgniteCheckedException {
assert desc != null;
assert msg != null;
if (log.isTraceEnabled())
log.trace("Sending message to Hadoop process [desc=" + desc + ", msg=" + msg + ']');
HadoopCommunicationClient client = null;
boolean closeOnRelease = true;
try {
client = reserveClient(desc);
client.sendMessage(desc, msg);
closeOnRelease = false;
} finally {
if (client != null) {
if (closeOnRelease) {
client.forceClose();
clients.remove(desc.processId(), client);
} else client.release();
}
}
}
/** {@inheritDoc} */
@Override
public void onDisconnected(GridNioSession ses, @Nullable Exception e) {
if (log.isDebugEnabled()) log.debug("Closed connection for session: " + ses);
if (e != null) U.error(log, "Session disconnected due to exception: " + ses, e);
HadoopProcessDescriptor desc = ses.meta(PROCESS_META);
if (desc != null) {
HadoopCommunicationClient rmv = clients.remove(desc.processId());
if (rmv != null) rmv.forceClose();
}
HadoopMessageListener lsnr0 = lsnr;
if (lsnr0 != null)
// Notify listener about connection close.
lsnr0.onConnectionLost(desc);
}
/**
* @param desc Process descriptor.
* @return Client.
* @throws IgniteCheckedException If failed.
*/
@Nullable
protected HadoopCommunicationClient createNioClient(HadoopProcessDescriptor desc)
throws IgniteCheckedException {
assert desc != null;
int shmemPort = desc.sharedMemoryPort();
// If remote node has shared memory server enabled and has the same set of MACs
// then we are likely to run on the same host and shared memory communication could be tried.
if (shmemPort != -1 && locProcDesc.parentNodeId().equals(desc.parentNodeId())) {
try {
return createShmemClient(desc, shmemPort);
} catch (IgniteCheckedException e) {
if (e.hasCause(IpcOutOfSystemResourcesException.class))
// Has cause or is itself the IpcOutOfSystemResourcesException.
LT.warn(log, null, OUT_OF_RESOURCES_TCP_MSG);
else if (log.isDebugEnabled())
log.debug(
"Failed to establish shared memory connection with local hadoop process: " + desc);
}
}
return createTcpClient(desc);
}
/**
* Returns existing or just created client to node.
*
* @param desc Node to which client should be open.
* @return The existing or just created client.
* @throws IgniteCheckedException Thrown if any exception occurs.
*/
private HadoopCommunicationClient reserveClient(HadoopProcessDescriptor desc)
throws IgniteCheckedException {
assert desc != null;
UUID procId = desc.processId();
while (true) {
HadoopCommunicationClient client = clients.get(procId);
if (client == null) {
if (log.isDebugEnabled())
log.debug(
"Did not find client for remote process [locProcDesc="
+ locProcDesc
+ ", desc="
+ desc
+ ']');
// Do not allow concurrent connects.
Object sync = locks.lock(procId);
try {
client = clients.get(procId);
if (client == null) {
HadoopCommunicationClient old = clients.put(procId, client = createNioClient(desc));
assert old == null;
}
} finally {
locks.unlock(procId, sync);
}
assert client != null;
}
if (client.reserve()) return client;
else
// Client has just been closed by idle worker. Help it and try again.
clients.remove(procId, client);
}
}
/**
* Establish TCP connection to remote hadoop process and returns client.
*
* @param desc Process descriptor.
* @return Client.
* @throws IgniteCheckedException If failed.
*/
protected HadoopCommunicationClient createTcpClient(HadoopProcessDescriptor desc)
throws IgniteCheckedException {
String addr = desc.address();
int port = desc.tcpPort();
if (log.isDebugEnabled())
log.debug(
"Trying to connect to remote process [locProcDesc="
+ locProcDesc
+ ", desc="
+ desc
+ ']');
boolean conn = false;
HadoopTcpNioCommunicationClient client = null;
IgniteCheckedException errs = null;
int connectAttempts = 1;
long connTimeout0 = connTimeout;
int attempt = 1;
while (!conn) { // Reconnection on handshake timeout.
try {
SocketChannel ch = SocketChannel.open();
ch.configureBlocking(true);
ch.socket().setTcpNoDelay(tcpNoDelay);
ch.socket().setKeepAlive(true);
if (sockRcvBuf > 0) ch.socket().setReceiveBufferSize(sockRcvBuf);
if (sockSndBuf > 0) ch.socket().setSendBufferSize(sockSndBuf);
ch.socket().connect(new InetSocketAddress(addr, port), (int) connTimeout);
HandshakeFinish fin = new HandshakeFinish();
GridNioSession ses = nioSrvr.createSession(ch, F.asMap(HANDSHAKE_FINISH_META, fin)).get();
client = new HadoopTcpNioCommunicationClient(ses);
if (log.isDebugEnabled()) log.debug("Waiting for handshake finish for client: " + client);
fin.await(connTimeout0);
conn = true;
} catch (HadoopHandshakeTimeoutException e) {
if (client != null) {
client.forceClose();
client = null;
}
if (log.isDebugEnabled())
log.debug(
"Handshake timedout (will retry with increased timeout) [timeout="
+ connTimeout0
+ ", desc="
+ desc
+ ", port="
+ port
+ ", err="
+ e
+ ']');
if (attempt == reconCnt || connTimeout0 > maxConnTimeout) {
if (log.isDebugEnabled())
log.debug(
"Handshake timed out (will stop attempts to perform the handshake) "
+ "[timeout="
+ connTimeout0
+ ", maxConnTimeout="
+ maxConnTimeout
+ ", attempt="
+ attempt
+ ", reconCnt="
+ reconCnt
+ ", err="
+ e.getMessage()
+ ", addr="
+ addr
+ ']');
if (errs == null)
errs =
new IgniteCheckedException(
"Failed to connect to remote Hadoop process "
+ "(is process still running?) [desc="
+ desc
+ ", addrs="
+ addr
+ ']');
errs.addSuppressed(e);
break;
} else {
attempt++;
connTimeout0 *= 2;
// Continue loop.
}
} catch (Exception e) {
if (client != null) {
client.forceClose();
client = null;
}
if (log.isDebugEnabled())
log.debug("Client creation failed [addr=" + addr + ", port=" + port + ", err=" + e + ']');
if (X.hasCause(e, SocketTimeoutException.class))
LT.warn(
log,
null,
"Connect timed out (consider increasing 'connTimeout' "
+ "configuration property) [addr="
+ addr
+ ", port="
+ port
+ ']');
if (errs == null)
errs =
new IgniteCheckedException(
"Failed to connect to remote Hadoop process (is process still running?) "
+ "[desc="
+ desc
+ ", addrs="
+ addr
+ ']');
errs.addSuppressed(e);
// Reconnect for the second time, if connection is not established.
if (connectAttempts < 2
&& (e instanceof ConnectException || X.hasCause(e, ConnectException.class))) {
connectAttempts++;
continue;
}
break;
}
}
if (client == null) {
assert errs != null;
if (X.hasCause(errs, ConnectException.class))
LT.warn(
log,
null,
"Failed to connect to a remote Hadoop process (is process still running?). "
+ "Make sure operating system firewall is disabled on local and remote host) "
+ "[addrs="
+ addr
+ ", port="
+ port
+ ']');
throw errs;
}
if (log.isDebugEnabled()) log.debug("Created client: " + client);
return client;
}
/**
* Creates new shared memory communication server.
*
* @return Server.
* @throws IgniteCheckedException If failed.
*/
@Nullable
private IpcSharedMemoryServerEndpoint resetShmemServer() throws IgniteCheckedException {
if (boundTcpShmemPort >= 0)
throw new IgniteCheckedException(
"Shared memory server was already created on port " + boundTcpShmemPort);
if (shmemPort == -1 || U.isWindows()) return null;
IgniteCheckedException lastEx = null;
// If configured TCP port is busy, find first available in range.
for (int port = shmemPort; port < shmemPort + locPortRange; port++) {
try {
IpcSharedMemoryServerEndpoint srv =
new IpcSharedMemoryServerEndpoint(
log.getLogger(IpcSharedMemoryServerEndpoint.class),
locProcDesc.processId(),
gridName);
srv.setPort(port);
srv.omitOutOfResourcesWarning(true);
srv.start();
boundTcpShmemPort = port;
// Ack Port the TCP server was bound to.
if (log.isInfoEnabled())
log.info(
"Successfully bound shared memory communication to TCP port [port="
+ boundTcpShmemPort
+ ", locHost="
+ locHost
+ ']');
return srv;
} catch (IgniteCheckedException e) {
lastEx = e;
if (log.isDebugEnabled())
log.debug(
"Failed to bind to local port (will try next port within range) [port="
+ port
+ ", locHost="
+ locHost
+ ']');
}
}
// If free port wasn't found.
throw new IgniteCheckedException(
"Failed to bind shared memory communication to any port within range [startPort="
+ locPort
+ ", portRange="
+ locPortRange
+ ", locHost="
+ locHost
+ ']',
lastEx);
}
/** {@inheritDoc} */
@Override
public void onMessageReceived(GridNioSession ses, Object msg) throws IgniteCheckedException {
HadoopProcessDescriptor desc = ses.meta(PROCESS_META);
UUID rmtProcId = desc == null ? null : desc.processId();
if (rmtProcId == null) {
if (!(msg instanceof ProcessHandshakeMessage)) {
log.warning(
"Invalid handshake message received, will close connection [ses="
+ ses
+ ", msg="
+ msg
+ ']');
ses.close();
return;
}
ProcessHandshakeMessage nId = (ProcessHandshakeMessage) msg;
if (log.isDebugEnabled())
log.debug("Received handshake message [ses=" + ses + ", msg=" + msg + ']');
ses.addMeta(PROCESS_META, nId.processDescriptor());
if (!ses.accepted())
// Send handshake reply.
ses.send(locIdMsg);
else {
//
rmtProcId = nId.processDescriptor().processId();
if (log.isDebugEnabled()) log.debug("Finished handshake with remote client: " + ses);
Object sync = locks.tryLock(rmtProcId);
if (sync != null) {
try {
if (clients.get(rmtProcId) == null) {
if (log.isDebugEnabled())
log.debug("Will reuse session for descriptor: " + rmtProcId);
// Handshake finished flag is true.
clients.put(rmtProcId, new HadoopTcpNioCommunicationClient(ses));
} else {
if (log.isDebugEnabled())
log.debug(
"Will not reuse client as another already exists [locProcDesc="
+ locProcDesc
+ ", desc="
+ desc
+ ']');
}
} finally {
locks.unlock(rmtProcId, sync);
}
} else {
if (log.isDebugEnabled())
log.debug(
"Concurrent connection is being established, will not reuse client session ["
+ "locProcDesc="
+ locProcDesc
+ ", desc="
+ desc
+ ']');
}
}
if (log.isDebugEnabled())
log.debug(
"Handshake is finished for session [ses="
+ ses
+ ", locProcDesc="
+ locProcDesc
+ ']');
HandshakeFinish to = ses.meta(HANDSHAKE_FINISH_META);
if (to != null) to.finish();
// Notify session opened (both parties).
proceedSessionOpened(ses);
} else {
if (msgQueueLimit > 0) {
GridNioMessageTracker tracker = ses.meta(TRACKER_META);
if (tracker == null) {
GridNioMessageTracker old =
ses.addMeta(TRACKER_META, tracker = new GridNioMessageTracker(ses, msgQueueLimit));
assert old == null;
}
tracker.onMessageReceived();
}
proceedMessageReceived(ses, msg);
}
}