Chapter 2. NIO extensions

2.1. Benefits and shortcomings of the non-blocking I/O model

Contrary to the popular belief, the performance of NIO in terms of raw data throughput is significantly lower than that of blocking I/O. NIO does not necessarily fit all use cases and should be used only where appropriate:

  • handling of thousands of connections, a significant number of which can be idle.

  • handling high latency connections.

  • request / response handling needs to be decoupled.

2.2. Differences from other NIO frameworks

Solves similar problems as other frameworks, but has certain distinct features:

  • minimalistic, optimized for data volume intensive protocols such as HTTP.

  • efficient memory management: data consumer can read only as much input data as it can process without having to allocate more memory.

  • direct access to the NIO channels where possible.

2.3. I/O reactor

HttpCore NIO is based on the Reactor pattern as described by Doug Lea. The purpose of I/O reactors is to react to I/O events and to dispatch event notifications to individual I/O sessions. The main idea of I/O reactor pattern is to break away from the one thread per connection model imposed by the classic blocking I/O model. The IOReactor interface represents an abstract object implementing the Reactor pattern. Internally, IOReactor implementations encapsulate functionality of the NIO java.nio.channels.Selector.

I/O reactors usually employ a small number of dispatch threads (often as few as one) to dispatch I/O event notifications to a much greater number (often as many as several thousands) of I/O sessions or connections. It is generally recommended to have one dispatch thread per CPU core.

HttpParams params = new BasicHttpParams();
int workerCount = 2;
IOReactor ioreactor = new DefaultConnectingIOReactor(workerCount, 
    params);

2.3.1. I/O dispatchers

IOReactor implementations make use of the IOEventDispatch interface to notify clients of events pending for a particular session. All methods of the IOEventDispatch are executed on a dispatch thread of the I/O reactor. Therefore, it is important that processing that takes place in the event methods will not block the dispatch thread for too long, as the I/O reactor will be unable to react to other events.

HttpParams params = new BasicHttpParams();
IOReactor ioreactor = new DefaultConnectingIOReactor(2, params);
     
IOEventDispatch eventDispatch = new MyIOEventDispatch();
ioreactor.execute(eventDispatch);

Generic I/O events as defined by the IOEventDispatch interface:

  • connected Triggered when a new session has been created.

  • inputReady Triggered when the session has pending input.

  • outputReady Triggered when the session is ready for output.

  • timeout Triggered when the session has timed out.

  • disconnected Triggered when the session has been terminated.

2.3.2. I/O reactor shutdown

The shutdown of I/O reactors is a complex process and may usually take a while to complete. I/O reactors will attempt to gracefully terminate all active I/O sessions and dispatch threads approximately within the specified grace period. If any of the I/O sessions fails to terminate correctly, the I/O reactor will forcibly shut down remaining sessions.

long gracePeriod = 3000L; // milliseconds
ioreactor.shutdown(gracePeriod);

The IOReactor#shutdown(long) method is safe to call from any thread.

2.3.3. I/O sessions

The IOSession interface represents a sequence of logically related data exchanges between two end points. IOSession encapsulates functionality of NIO java.nio.channels.SelectionKey and java.nio.channels.SocketChannel. The channel associated with the IOSession can be used to read data from and write data to the session.

IOSession iosession;
ReadableByteChannel ch = (ReadableByteChannel) iosession.channel();
ByteBuffer dst = ByteBuffer.allocate(2048); 
ch.read(dst);

2.3.4. I/O session state management

I/O sessions are not bound to an execution thread, therefore one cannot use the context of the thread to store a session's state. All details about a particular session must be stored within the session itself.

IOSession iosession;
Object someState;
iosession.setAttribute("state", someState);
Object currentState = iosession.getAttribute("state");

Please note that if several sessions make use of shared objects, access to those objects must be made thread-safe.

2.3.5. I/O session event mask

One can declare an interest in a particular type of I/O events for a particular I/O session by setting its event mask.

IOSession iosession;
iosession.setEventMask(SelectionKey.OP_READ | SelectionKey.OP_WRITE);

One can also toggle OP_READ and OP_WRITE flags individually.

iosession.setEvent(SelectionKey.OP_READ);
iosession.clearEvent(SelectionKey.OP_READ);

Event notifications will not take place if the corresponding interest flag is not set.

2.3.6. I/O session buffers

Quite often I/O sessions need to maintain internal I/O buffers in order to transform input / output data prior to returning it to the consumer or writing it to the underlying channel. Memory management in HttpCore NIO is based on the fundamental principle that the data consumer can read only as much input data as it can process without having to allocate more memory. That means, quite often some input data may remain unread in one of the internal or external session buffers. The I/O reactor can query the status of these session buffers, and make sure the consumer gets notified correctly as more data gets stored in one of the session buffers, thus allowing the consumer to read the remaining data once it is able to process it. I/O sessions can be made aware of the status of external session buffers using the SessionBufferStatus interface.

IOSession iosession;
SessionBufferStatus myBufferStatus = new MySessionBufferStatus(); 
iosession.setBufferStatus(myBufferStatus);
iosession.hasBufferedInput();
iosession.hasBufferedOutput();

2.3.7. I/O session shutdown

One can close an I/O session gracefully by calling IOSession#close() allowing the session to be closed in an orderly manner or by calling IOSession#shutdown() to forcibly close the underlying channel. The distinction between two methods is of primary importance for those types of I/O sessions that involve some sort of a session termination handshake such as SSL/TLS connections.

2.3.8. Listening I/O reactors

ListeningIOReactor represents an I/O reactor capable of listening for incoming connections on one or several ports.

ListeningIOReactor ioreactor;
        
ListenerEndpoint ep1 = ioreactor.listen(new InetSocketAddress(8081));
ListenerEndpoint ep2 = ioreactor.listen(new InetSocketAddress(8082));
ListenerEndpoint ep3 = ioreactor.listen(new InetSocketAddress(8083));

// Wait until all endpoints are up
ep1.waitFor();
ep2.waitFor();
ep3.waitFor();

Once an endpoint is fully initialized it starts accepting incoming connections and propagates I/O activity notifications to the IOEventDispatch instance.

One can obtain a set of registered endpoints at runtime, query the status of an endpoint at runtime, and close it if desired.

ListeningIOReactor ioreactor;
        
Set<ListenerEndpoint> eps = ioreactor.getEndpoints();
for (ListenerEndpoint ep: eps) {
    // Still active?
    System.out.println(ep.getAddress());
    if (ep.isClosed()) {
        // If not, has it terminated due to an exception?
        if (ep.getException() != null) {
            ep.getException().printStackTrace();
        }
    } else {
        ep.close();
    }
}

2.3.9. Connecting I/O reactors

ConnectingIOReactor represents an I/O reactor capable of establishing connections with remote hosts.

ConnectingIOReactor ioreactor;

SessionRequest sessionRequest = ioreactor.connect(
        new InetSocketAddress("www.google.com", 80), 
        null, null, null);

Opening a connection to a remote host usually tends to be a time consuming process and may take a while to complete. One can monitor and control the process of session initialization by means of the SessionRequestinterface.

// Make sure the request times out if connection 
// has not been established after 1 sec
sessionRequest.setConnectTimeout(1000);
// Wait for the request to complete
sessionRequest.waitFor();
// Has request terminated due to an exception?
if (sessionRequest.getException() != null) {
    sessionRequest.getException().printStackTrace();
}
// Get hold of the new I/O session
IOSession iosession = sessionRequest.getSession();

SessionRequest implementations are expected to be thread-safe. Session request can be aborted at any time by calling IOSession#cancel() from another thread of execution.

if (!sessionRequest.isCompleted()) {
    sessionRequest.cancel();
}

One can pass several optional parameters to the ConnectingIOReactor#connect() method to exert a greater control over the process of session initialization.

A non-null local socket address parameter can be used to bind the socket to a specific local address.

ConnectingIOReactor ioreactor;

SessionRequest sessionRequest = ioreactor.connect(
        new InetSocketAddress("www.google.com", 80), 
        new InetSocketAddress("192.168.0.10", 1234), 
        null, null);

One can provide an attachment object, which will be added to the new session's context upon initialization. This object can be used to pass an initial processing state to the protocol handler.

SessionRequest sessionRequest = ioreactor.connect(
        new InetSocketAddress("www.google.com", 80), 
        null, new HttpHost("www.google.ru"), null);
        
IOSession iosession = sessionRequest.getSession();
HttpHost virtualHost = (HttpHost) iosession.getAttribute(
    IOSession.ATTACHMENT_KEY);

It is often desirable to be able to react to the completion of a session request asynchronously without having to wait for it, blocking the current thread of execution. One can optionally provide an implementation SessionRequestCallback interface to get notified of events related to session requests, such as request completion, cancellation, failure or timeout.

ConnectingIOReactor ioreactor;

SessionRequest sessionRequest = ioreactor.connect(
        new InetSocketAddress("www.google.com", 80), null, null,
        new SessionRequestCallback() {

            public void cancelled(SessionRequest request) {
            }

            public void completed(SessionRequest request) {
                System.out.println("new connection to " + 
                    request.getRemoteAddress());
            }

            public void failed(SessionRequest request) {
                if (request.getException() != null) {
                    request.getException().printStackTrace();
                }
            }

            public void timeout(SessionRequest request) {
            }
            
        });

2.3.10. Queuing of I/O interest set operations

Several older JRE implementations (primarily from IBM) include what Java API documentation refers to as a naive implementation of the java.nio.channels.SelectionKey class. The problem with java.nio.channels.SelectionKey in such JREs is that reading or writing of the I/O interest set may block indefinitely if the I/O selector is in the process of executing a select operation. HttpCore NIO can be configured to operate in a special mode wherein I/O interest set operations are queued and executed by on the dispatch thread only when the I/O selector is not engaged in a select operation.

HttpParams params = new BasicHttpParams();
NIOReactorParams.setInterestOpsQueueing(params, true);
ListeningIOReactor ioreactor = new DefaultListeningIOReactor(2, params);

2.4. I/O reactor exception handling

Protocol specific exceptions as well as those I/O exceptions thrown in the course of interaction with the session's channel are to be expected are to be dealt with by specific protocol handlers. These exceptions may result in termination of an individual session but should not affect the I/O reactor and all other active sessions. There are situations, however, when the I/O reactor itself encounters an internal problem such as an I/O exception in the underlying NIO classes or an unhandled runtime exception. Those types of exceptions are usually fatal and will cause the I/O reactor to shut down automatically.

There is a possibility to override this behaviour and prevent I/O reactors from shutting down automatically in case of a runtime exception or an I/O exception in internal classes. This can be accomplished by providing a custom implementation of the IOReactorExceptionHandler interface.

DefaultConnectingIOReactor ioreactor;

ioreactor.setExceptionHandler(new IOReactorExceptionHandler() {

    public boolean handle(IOException ex) {
        if (ex instanceof BindException) {
            // bind failures considered OK to ignore
            return true;
        }
        return false;
    }

    public boolean handle(RuntimeException ex) {
        if (ex instanceof UnsupportedOperationException) {
            // Unsupported operations considered OK to ignore
            return true;
        }
        return false;
    }
    
});

One needs to be very careful about discarding exceptions indiscriminately. It is often much better to let the I/O reactor shut down itself cleanly and restart it rather than leaving it in an inconsistent or unstable state.

2.4.1. I/O reactor audit log

If an I/O reactor is unable to automatically recover from an I/O or a runtime exception it will enter the shutdown mode. First off, it will close all active listeners and cancel all pending new session requests. Then it will attempt to close all active I/O sessions gracefully giving them some time to flush pending output data and terminate cleanly. Lastly, it will forcibly shut down those I/O sessions that still remain active after the grace period. This is a fairly complex process, where many things can fail at the same time and many different exceptions can be thrown in the course of the shutdown process. The I/O reactor will record all exceptions thrown during the shutdown process, including the original one that actually caused the shutdown in the first place, in an audit log. One can examine the audit log and decide whether it is safe to restart the I/O reactor.

DefaultConnectingIOReactor ioreactor;

// Give it 5 sec grace period
ioreactor.shutdown(5000);
List<ExceptionEvent> events = ioreactor.getAuditLog();
for (ExceptionEvent event: events) {
    System.err.println("Time: " + event.getTimestamp());
    event.getCause().printStackTrace();
}

2.5. Non-blocking HTTP connections

Effectively non-blocking HTTP connections are wrappers around IOSession with HTTP specific functionality. Non-blocking HTTP connections are stateful and not thread-safe. Input / output operations on non-blocking HTTP connections should be restricted to the dispatch events triggered by the I/O event dispatch thread.

2.5.1. Execution context of non-blocking HTTP connections

Non-blocking HTTP connections are not bound to a particular thread of execution and therefore they need to maintain their own execution context. Each non-blocking HTTP connection has an HttpContext instance associated with it, which can be used to maintain a processing state. The HttpContext instance is thread-safe and can be manipulated from multiple threads.

// Get non-blocking HTTP connection
DefaultNHttpClientConnection conn;
// State
Object myStateObject;

HttpContext context = conn.getContext();
context.setAttribute("state", myStateObject);

2.5.2. Working with non-blocking HTTP connections

At any point of time one can obtain the request and response objects currently being transferred over the non-blocking HTTP connection. Any of these objects, or both, can be null if there is no incoming or outgoing message currently being transferred.

NHttpConnection conn;

HttpRequest request = conn.getHttpRequest();
if (request != null) {
    System.out.println("Transferring request: " + 
        request.getRequestLine());
}
HttpResponse response = conn.getHttpResponse();
if (response != null) {
    System.out.println("Transferring response: " + 
        response.getStatusLine());
}

However, please note that the current request and the current response may not necessarily represent the same message exchange! Non-blocking HTTP connections can operate in a full duplex mode. One can process incoming and outgoing messages completely independently from one another. This makes non-blocking HTTP connections fully pipelining capable, but at same time implies that this is the job of the protocol handler to match logically related request and the response messages.

Over-simplified process of submitting a request on the client side may look like this:

// Obtain HTTP connection
NHttpClientConnection conn;

// Obtain execution context
HttpContext context = conn.getContext();

// Obtain processing state
Object state = context.getAttribute("state");

// Generate a request based on the state information
HttpRequest request = new BasicHttpRequest("GET", "/");
    
conn.submitRequest(request);
System.out.println(conn.isRequestSubmitted());

Over-simplified process of submitting a response on the server side may look like this:

// Obtain HTTP connection
NHttpServerConnection conn;

// Obtain execution context
HttpContext context = conn.getContext();

// Obtain processing state
Object state = context.getAttribute("state");

// Generate a response based on the state information
HttpResponse response = new BasicHttpResponse(HttpVersion.HTTP_1_1, 
    HttpStatus.SC_OK, "OK");
BasicHttpEntity entity = new BasicHttpEntity();
entity.setContentType("text/plain");
entity.setChunked(true);
response.setEntity(entity);
    
conn.submitResponse(response);
System.out.println(conn.isResponseSubmitted());

Please note that one should rarely need to transmit messages using these low level methods and should use appropriate higher level HTTP service implementations instead.

2.5.3. HTTP I/O control

All non-blocking HTTP connections classes implement IOControl interface, which represents a subset of connection functionality for controlling interest in I/O even notifications. IOControl instances are expected to be fully thread-safe. Therefore IOControl can be used to request / suspend I/O event notifications from any thread.

One must take special precautions when interacting with non-blocking connections. HttpRequest and HttpResponse are not thread-safe. It is generally advisable that all input / output operations on a non-blocking connection are executed from the I/O event dispatch thread.

The following pattern is recommended:

  • Use IOControl interface to pass control over connection's I/O events to another thread / session.

  • If input / output operations need be executed on that particular connection, store all the required information (state) in the connection context and request the appropriate I/O operation by calling IOControl#requestInput() or IOControl#requestOutput() method.

  • Execute the required operations from the event method on the dispatch thread using information stored in connection context.

Please note all operations that take place in the event methods should not block for too long, because while the dispatch thread remains blocked in one session, it is unable to process events for all other sessions. I/O operations with the underlying channel of the session are not a problem as they are guaranteed to be non-blocking.

2.5.4. Non-blocking content transfer

The process of content transfer for non-blocking connections works completely differently compared to that of blocking connections, as non-blocking connections need to accommodate to the asynchronous nature of the NIO model. The main distinction between two types of connections is inability to use the usual, but inherently blocking java.io.InputStream and java.io.OutputStream classes to represent streams of inbound and outbound content. HttpCore NIO provides ContentEncoder and ContentDecoder interfaces to handle the process of asynchronous content transfer. Non-blocking HTTP connections will instantiate the appropriate implementation of a content codec based on properties of the entity enclosed with the message.

Non-blocking HTTP connections will fire input events until the content entity is fully transferred.

//Obtain content decoder
ContentDecoder decoder;
//Read data in
ByteBuffer dst = ByteBuffer.allocate(2048); 
decoder.read(dst);
// Decode will be marked as complete when 
// the content entity is fully transferred
if (decoder.isCompleted()) {
    // Done
}

Non-blocking HTTP connections will fire output events until the content entity is marked as fully transferred.

// Obtain content encoder
ContentEncoder encoder;
// Prepare output data 
ByteBuffer src = ByteBuffer.allocate(2048); 
// Write data out
encoder.write(src);
// Mark content entity as fully transferred when done
encoder.complete();
}

Please note, one still has to provide an !HttpEntity instance when submitting an entity enclosing message to the non-blocking HTTP connection. Properties of that entity will be used to initialize an ContentEncoder instance to be used for transferring entity content. Non-blocking HTTP connections, however, ignore inherently blocking HttpEntity#getContent() and HttpEntity#writeTo() methods of the enclosed entities.

// Obtain HTTP connection
NHttpServerConnection conn;

HttpResponse response = new BasicHttpResponse(HttpVersion.HTTP_1_1, 
    HttpStatus.SC_OK, "OK");
BasicHttpEntity entity = new BasicHttpEntity();
entity.setContentType("text/plain");
entity.setChunked(true);
entity.setContent(null);
response.setEntity(entity);
    
conn.submitResponse(response);

Likewise, incoming entity enclosing message will have an HttpEntity instance associated with them, but an attempt to call HttpEntity#getContent() or HttpEntity#writeTo() methods will cause an java.lang.IllegalStateException. The HttpEntity instance can be used to determine properties of the incoming entity such as content length.

// Obtain HTTP connection
NHttpClientConnection conn;

HttpResponse response = conn.getHttpResponse();
HttpEntity entity = response.getEntity();
if (entity != null) {
    System.out.println(entity.getContentType());
    System.out.println(entity.getContentLength());
    System.out.println(entity.isChunked());
}

2.5.5. Supported non-blocking content transfer mechanisms

Default implementations of the non-blocking HTTP connection interfaces support three content transfer mechanisms defined by the HTTP/1.1 specification:

  • Content-Length delimited:  The end of the content entity is determined by the value of the Content-Length header. Maximum entity length: Long#MAX_VALUE.

  • Identity coding:  The end of the content entity is demarcated by closing the underlying connection (end of stream condition). For obvious reasons the identity encoding can only be used on the server side. Max entity length: unlimited.

  • Chunk coding:  The content is sent in small chunks. Max entity length: unlimited.

The appropriate content codec will be created automatically depending on properties of the entity enclosed with the message.

2.5.6. Direct channel I/O

Content codes are optimized to read data directly from or write data directly to the underlying I/O session's channel, whenever possible avoiding intermediate buffering in a session buffer. Moreover, those codecs that do not perform any content transformation such as Content-Length delimited and identity can leverage NIO java.nio.FileChannel methods for significantly improved performance of file transfer operations both inbound and outbound.

If the actual content decoder implements FileContentDecoder one can make use of its methods to read incoming content directly to a file bypassing an intermediate java.nio.ByteBuffer.

//Obtain content decoder
ContentDecoder decoder;
//Prepare file channel
FileChannel dst;
//Make use of direct file I/O if possible
if (decoder instanceof FileContentDecoder) {
    long Bytesread = ((FileContentDecoder) decoder)
        .transfer(dst, 0, 2048);
     // Decode will be marked as complete when 
     // the content entity is fully transmitted
     if (decoder.isCompleted()) {
         // Done
     }
}

If the actual content encoder implements FileContentEncoder one can make use of its methods to write outgoing content directly from a file bypassing an intermediate java.nio.ByteBuffer.

// Obtain content encoder
ContentEncoder encoder;
// Prepare file channel
FileChannel src;
// Make use of direct file I/O if possible
if (encoder instanceof FileContentEncoder) {
    // Write data out
    long bytesWritten = ((FileContentEncoder) encoder)
        .transfer(src, 0, 2048); 
    // Mark content entity as fully transferred when done
    encoder.complete();
}

2.6. HTTP I/O event dispatchers

HTTP I/O event dispatchers serve to convert generic I/O events triggered by an I/O reactor to HTTP protocol specific events. They rely on NHttpClientHandler and NHttpServiceHandler interfaces to propagate HTTP protocol events to a HTTP protocol handler.

Server side HTTP I/O events as defined by the NHttpServiceHandler interface:

  • connected Triggered when a when a new incoming connection has been created.

  • requestReceived Triggered when a new HTTP request is received. The connection passed as a parameter to this method is guaranteed to return a valid HTTP request object. If the request received encloses a request entity this method will be followed a series of inputReady events to transfer the request content.

  • inputReady Triggered when the underlying channel is ready for reading a new portion of the request entity through the corresponding content decoder. If the content consumer is unable to process the incoming content, input event notifications can be temporarily suspended using IOControl interface.

  • responseReady Triggered when the connection is ready to accept new HTTP response. The protocol handler does not have to submit a response if it is not ready.

  • outputReady Triggered when the underlying channel is ready for writing a next portion of the response entity through the corresponding content encoder. If the content producer is unable to generate the outgoing content, output event notifications can be temporarily suspended using IOControl interface.

  • exception Triggered when an I/O error occurrs while reading from or writing to the underlying channel or when an HTTP protocol violation occurs while receiving an HTTP request.

  • exception Triggered when an I/O error occurrs while reading from or writing to the underlying channel or when an HTTP protocol violation occurs while receiving an HTTP request.

  • timeout Triggered when no input is detected on this connection over the maximum period of inactivity.

  • closed Triggered when the connection has been closed.

Client side HTTP I/O events as defined by the NHttpClientHandler interface:

  • connected Triggered when a new outgoing connection has been created. The attachment object passed as a parameter to this event is an arbitrary object that was attached to the session request.

  • requestReady Triggered when the connection is ready to accept new HTTP request. The protocol handler does not have to submit a request if it is not ready.

  • outputReady Triggered when the underlying channel is ready for writing a next portion of the request entity through the corresponding content encoder. If the content producer is unable to generate the outgoing content, output event notifications can be temporarily suspended using IOControl interface.

  • responseReceived Triggered when an HTTP response is received. The connection passed as a parameter to this method is guaranteed to return a valid HTTP response object. If the response received encloses a response entity this method will be followed a series of inputReady events to transfer the response content.

  • inputReady Triggered when the underlying channel is ready for reading a new portion of the response entity through the corresponding content decoder. If the content consumer is unable to process the incoming content, input event notifications can be temporarily suspended using IOControl interface.

  • exception Triggered when an I/O error occurs while reading from or writing to the underlying channel or when an HTTP protocol violation occurs while receiving an HTTP response..

  • exception Triggered when an I/O error occurrs while reading from or writing to the underlying channel or when an HTTP protocol violation occurs while receiving an HTTP request.

  • timeout Triggered when no input is detected on this connection over the maximum period of inactivity.

  • closed Triggered when the connection has been closed.

2.7. Non-blocking HTTP entities

As discussed previously the process of content transfer for non-blocking connections works completely differently compared to that for blocking connections. For obvious reasons classic I/O abstraction based on inherently blocking java.io.InputStream and java.io.OutputStream classes is not applicable to the asynchronous process of data transfer. Therefore, non-blocking HTTP entities provide NIO specific extensions to the !HttpEntity interface: ProducingNHttpEntity and ConsumingNHttpEntity interfaces. Implementation classes of these interfaces may throw java.lang.UnsupportedOperationException from HttpEntity#getContent() or HttpEntity#writeTo() if a particular implementation is unable to represent its content stream as instance of java.io.InputStream or cannot stream its content out to an java.io.OutputStream.

2.7.1. Content consuming non-blocking HTTP entity

ConsumingNHttpEntity interface represents a non-blocking entity that allows content to be consumed from a content decoder. ConsumingNHttpEntity extends the base HttpEntity interface with a number of NIO specific notification methods:

  • consumeContent Notification that content is available to be read from the decoder. IOControl instance passed as a parameter to the method can be used to suspend input events if the entity is temporarily unable to allocate more storage to accommodate all incoming content.

  • finish Notification that any resources allocated for reading can be released.

The following implementations of ConsumingNHttpEntity provided by HttpCore NIO:

2.7.1.1. BufferingNHttpEntity

BufferingNHttpEntity is a subclass of HttpEntityWrapper that consumes all incoming content into memory. Once the content body has been fully received it can be retrieved as an java.io.InputStream via HttpEntity#getContent() , or written to an output stream via HttpEntity#writeTo() .

2.7.1.2. ConsumingNHttpEntityTemplate

ConsumingNHttpEntityTemplate is a subclass of HttpEntityWrapper that that decorates the incoming HTTP entity and delegates the handling of incoming content to a ContentListener instance.

static class FileWriteListener implements ContentListener {

    private final FileChannel fileChannel;
    private long idx = 0;

    public FileWriteListener(File file) throws IOException {
        this.fileChannel = new FileInputStream(file).getChannel();
    }

    public void contentAvailable(
    ContentDecoder decoder, IOControl ioctrl) throws IOException {
        long transferred;
        if (decoder instanceof FileContentDecoder) {
            transferred = ((FileContentDecoder) decoder).transfer(
                    fileChannel, idx, Long.MAX_VALUE);
        } else {
            transferred = fileChannel.transferFrom(
                    new ContentDecoderChannel(decoder), 
                    idx, Long.MAX_VALUE);
        }
        if (transferred > 0) {
            idx += transferred;
        }
    }

    public void finished() {
        try {
            fileChannel.close();
        } catch(IOException ignored) {}
    }
    
}

HttpEntity incomingEntity;

File file = new File("buffer.bin");
ConsumingNHttpEntity entity = new ConsumingNHttpEntityTemplate(
        incomingEntity, 
        new FileWriteListener(file)); 

2.7.2. Content producing non-blocking HTTP entity

ProducingNHttpEntity interface represents a non-blocking entity that allows content to be written to a content encoder. ProducingNHttpEntity extends the base HttpEntity interface with a number of NIO specific notification methods:

  • produceContent Notification that content can be written to the encoder. IOControl instance passed as a parameter to the method can be used to temporarily suspend output events if the entity is unable to produce more content. Please note one must call ContentEncoder#complete() to inform the underlying connection that all content has been written. Failure to do so could result in the entity never being correctly delimited.

  • finish Notification that any resources allocated for writing can be released.

The following implementations of ProducingNHttpEntity provided by HttpCore NIO:

2.7.2.1. NByteArrayEntity

This is a simple self contained repeatable entity, which receives its content from a given byte array. This byte array is supplied to the constructor.

String myData = "Hello world on the other side!!";
NByteArrayEntity entity = new NByteArrayEntity(myData.getBytes()); 

2.7.2.2. NStringEntity

It's is a simple, self contained, repeatable entity that retrieves its data from a java.lang.String object. It has 2 constructors, one simply constructs with a given string where the other also takes a character encoding for the data in the java.lang.String.

String myData = "Hello world on the other side!!";
// construct without a character encoding
NStringEntity myEntity1 = new NStringEntity(myData);
// alternatively construct with an encoding
NStringEntity myEntity2 = new NStringEntity(myData, "UTF-8");

2.7.2.3. NFileEntity

This entity reads its content body from a file. This class is mostly used to stream large files of different types, so one needs to supply the content type of the file to make sure the content can be correctly recognized and processed by the recipient.

File staticFile = new File("/path/to/myapp.jar");
NHttpEntity entity = new NFileEntity(staticFile, 
    "application/java-archive");

The NHttpEntity will make use of the direct channel I/O whenever possible, provided the content encoder is capable of transferring data directly from a file to the socket of the underlying connection.

2.8. Non-blocking HTTP protocol handlers

2.8.1. Asynchronous HTTP service handler

AsyncNHttpServiceHandler is a fully asynchronous HTTP server side protocol handler that implements the essential requirements of the HTTP protocol for the server side message processing as described by RFC 2616. AsyncNHttpServiceHandler is capable of processing HTTP requests with nearly constant memory footprint for individual HTTP connections. The handler stores headers of HTTP messages in memory, while content of message bodies is streamed directly from the entity to the underlying channel (and vice versa) using ConsumingNHttpEntity and ProducingNHttpEntity interfaces.

When using this implementation, it is important to ensure that entities supplied for writing implement ProducingNHttpEntity. Doing so will allow the entity to be written out asynchronously. If entities supplied for writing do not implement the ProducingNHttpEntity interface, a delegate is added that buffers the entire contents in memory. Additionally, the buffering might take place in the I/O dispatch thread, which could cause I/O to block temporarily. For best results, one must ensure that all entities set on HTTP responses from NHttpRequestHandler implement ProducingNHttpEntity.

If incoming requests enclose a content entity, NHttpRequestHandler instances are expected to return a ConsumingNHttpEntity for reading the content. After the entity is finished reading the data, NHttpRequestHandler#handle() method is called to generate a response.

AsyncNHttpServiceHandler relies on HttpProcessor to generate mandatory protocol headers for all outgoing messages and apply common, cross-cutting message transformations to all incoming and outgoing messages, whereas individual HTTP request handlers are expected to take care of application specific content generation and processing.

HttpParams params;
// Initialize HTTP parameters
HttpProcessor httpproc;
// Initialize HTTP processor

AsyncNHttpServiceHandler handler = new AsyncNHttpServiceHandler(
        httpproc,
        new DefaultHttpResponseFactory(),
        new DefaultConnectionReuseStrategy(),
        params);

2.8.1.1. Non-blocking HTTP request handlers

NHttpRequestHandler interface represents a routine for processing of a specific group of non-blocking HTTP requests. NHttpRequestHandler implementations are expected to take care of protocol specific aspects, whereas individual request handlers are expected take care of application specific HTTP processing. The main purpose of a request handler is to generate a response object with a content entity to be send back to the client in response to the given request.

NHttpRequestHandler myRequestHandler = new NHttpRequestHandler() {

    public ConsumingNHttpEntity entityRequest(
            HttpEntityEnclosingRequest request, 
            HttpContext context) throws HttpException, IOException {
        // Buffer imcoming content in memory for simplicity 
        return new BufferingNHttpEntity(request.getEntity(),
                new HeapByteBufferAllocator());
    }

    public void handle(
            HttpRequest request, 
            HttpResponse response,
            NHttpResponseTrigger trigger,
            HttpContext context) throws HttpException, IOException {
        response.setStatusCode(HttpStatus.SC_OK);
        response.addHeader("Content-Type", "text/plain");
        response.setEntity(
            new NStringEntity("some important message"));
        // Submit response immediately for simplicity
        trigger.submitResponse(response);
    }
    
};

Request handlers must be implemented in a thread-safe manner. Similarly to servlets, request handlers should not use instance variables unless access to those variables are synchronized.

2.8.1.2. Asynchronous response trigger

The most fundamental difference of the non-blocking request handlers compared to their blocking counterparts is ability to defer transmission of the HTTP response back to the client without blocking the I/O thread by delegating the process of handling the HTTP request to a worker thread. The worker thread can use the instance of NHttpResponseTrigger passed as a parameter to the NHttpRequestHandler#handle method to submit a response as at a later point of time once the response becomes available.

NHttpRequestHandler myRequestHandler = new NHttpRequestHandler() {

    public ConsumingNHttpEntity entityRequest(
            HttpEntityEnclosingRequest request, 
            HttpContext context) throws HttpException, IOException {
        // Buffer imcoming content in memory for simplicity 
        return new BufferingNHttpEntity(request.getEntity(),
                new HeapByteBufferAllocator());
    }

    public void handle(
            HttpRequest request, 
            HttpResponse response,
            NHttpResponseTrigger trigger,
            HttpContext context) 
                throws HttpException, IOException {
        new Thread() {
            
        @Override
        public void run() {
            try { 
                Thread.sleep(10); 
            } 
            catch(InterruptedException ie) {}
            try {
                URI uri = new URI(request.getRequestLine().getUri());
                response.setStatusCode(HttpStatus.SC_OK);
                response.addHeader("Content-Type", "text/plain");
                response.setEntity(
                    new NStringEntity("some important message"));
                trigger.submitResponse(response);                    
            } catch(URISyntaxException ex) {
                trigger.handleException(
                    new HttpException("Invalid request URI: " + 
                    ex.getInput()));
            }
        }
            
        }.start();
    }
    
};

Please note HttpResponse objects are not thread-safe and may not be modified concurrently. Non-blocking request handlers must ensure the HTTP response cannot be accessed by more than one thread at a time.

2.8.1.3. Non-blocking request handler resolver

The management of non-blocking HTTP request handlers is quite similar to that of blocking HTTP request handlers. Usually an instance of NHttpRequestHandlerResolver is used to maintain a registry of request handlers and to matches a request URI to a particular request handler. HttpCore includes only a very simple implementation of the request handler resolver based on a trivial pattern matching algorithm: NHttpRequestHandlerRegistry supports only three formats: *, <uri>* and *<uri>.

// Initialize asynchronous protocol handler
AsyncNHttpServiceHandler handler;

NHttpRequestHandlerRegistry handlerResolver = 
    new NHttpRequestHandlerRegistry();
handlerReqistry.register("/service/*", myRequestHandler1);
handlerReqistry.register("*.do", myRequestHandler2);
handlerReqistry.register("*", myRequestHandler3);

handler.setHandlerResolver(handlerResolver);

Users are encouraged to provide more sophisticated implementations of NHttpRequestHandlerResolver, for instance, based on regular expressions.

2.8.2. Asynchronous HTTP client handler

AsyncNHttpClientHandler is a fully asynchronous HTTP client side protocol handler that implements the essential requirements of the HTTP protocol for the client side message processing as described by RFC 2616. AsyncNHttpClientHandler is capable of executing HTTP requests with nearly constant memory footprint for individual HTTP connections. The handler stores headers of HTTP messages in memory, while content of message bodies is streamed directly from the entity to the underlying channel (and vice versa) using ConsumingNHttpEntity and ProducingNHttpEntity interfaces.

When using this implementation, it is important to ensure that entities supplied for writing implement ProducingNHttpEntity. Doing so will allow the entity to be written out asynchronously. If entities supplied for writing do not implement the ProducingNHttpEntity interface, a delegate is added that buffers the entire contents in memory. Additionally, the buffering might take place in the I/O dispatch thread, which could cause I/O to block temporarily. For best results, one must ensure that all entities set on HTTP requests from NHttpRequestExecutionHandler implement ProducingNHttpEntity.

If incoming responses enclose a content entity, NHttpRequestExecutionHandler is expected to return a ConsumingNHttpEntity for reading the content. After the entity is finished reading the data, NHttpRequestExecutionHandler#handleResponse() method is called to process the response.

If incoming responses enclose a content entity, NHttpRequestExecutionHandler is expected to return a ConsumingNHttpEntity for reading the content. After the entity is finished reading the data, NHttpRequestExecutionHandler#handleResponse() method is called to process the response.

AsyncNHttpClientHandler relies on HttpProcessor to generate mandatory protocol headers for all outgoing messages and apply common, cross-cutting message transformations to all incoming and outgoing messages, whereas HTTP request executor is expected to take care of application specific content generation and processing.

//  Initialize HTTP parameters
HttpParams params;
//Initialize HTTP processor
HttpProcessor httpproc;
//Create HTTP request execution handler
NHttpRequestExecutionHandler execHandler;

AsyncNHttpClientHandler handler = new AsyncNHttpClientHandler(
         httpproc,
         execHandler,
         new DefaultConnectionReuseStrategy(),
         params);

2.8.2.1. Asynchronous HTTP request execution handler

Asynchronous HTTP request execution handler can be used by client-side protocol handlers to trigger the submission of a new HTTP request and the processing of an HTTP response.

HTTP request execution exents as defined by the NHttpRequestExecutionHandler interface:

  • initalizeContext Triggered when a new connection has been established and the HTTP context needs to be initialized. The attachment object passed to this method is the same object which was passed to the connecting I/O reactor when the connection request was made. The attachment may optionally contain some state information required in order to correctly initialize the HTTP context.

  • submitRequest Triggered when the underlying connection is ready to send a new HTTP request to the target host. This method may return null if the client is not yet ready to send a request. In this case the connection will remain open and can be activated at a later point. If the request encloses an entity, the entity must be an instance of ProducingNHttpEntity.

  • responseEntity Triggered when a response is received with an entity. This method should return a ConsumingNHttpEntity that will be used to consume the entity. Null is a valid response value, and will indicate that the entity should be silently ignored. After the entity is fully consumed, handleResponse method is called to notify a full response and enclosed entity are ready to be processed.

  • handleResponse Triggered when an HTTP response is ready to be processed.

  • finalizeContext Triggered when the connection is terminated. This event can be used to release objects stored in the context or perform some other kind of cleanup.

NHttpRequestExecutionHandler execHandler = 
    new NHttpRequestExecutionHandler() {

    private final static String DONE_FLAG = "done";
    
    public void initalizeContext(
        HttpContext context, 
        Object attachment) {
        if (attachment != null) {
            HttpHost virtualHost = (HttpHost) attachment;
            context.setAttribute(ExecutionContext.HTTP_TARGET_HOST, 
                virtualHost);
        }
    }

    public void finalizeContext(HttpContext context) {
        context.removeAttribute(DONE_FLAG);
    }

    public HttpRequest submitRequest(HttpContext context) {
        // Submit HTTP GET once
        Object done = context.getAttribute(DONE_FLAG);
        if (done == null) {
            context.setAttribute(DONE_FLAG, Boolean.TRUE);
            return new BasicHttpRequest("GET", "/");
        } else {
            return null;
        }
    }
    
    public ConsumingNHttpEntity responseEntity(
        HttpResponse response, 
        HttpContext context) throws IOException {
        // Buffer imcoming content in memory for simplicity 
        return new BufferingNHttpEntity(response.getEntity(),
                new HeapByteBufferAllocator());
    }

    public void handleResponse(
        HttpResponse response, 
        HttpContext context) throws IOException {
        System.out.println(response.getStatusLine());
        if (response.getEntity() != null) {
            System.out.println(
                EntityUtils.toString(response.getEntity()));
        }
    }

};

2.8.3. Compatibility with blocking I/O

In addition to asynchronous protocol handlers described above HttpCore ships two variants of HTTP protocol handlers that emulate blocking I/O model on top of non-blocking one and allow message content to be produced and consumed using standard java.io.OutputStream / java.io.InputStream API. Compatibility protocol handlers can work with HTTP request handlers and request executors that rely on blocking HttpEntity implementations.

Compatibility protocol handlers rely on !HttpProcessor to generate mandatory protocol headers for all outgoing messages and apply common, cross-cutting message transformations to all incoming and outgoing messages, whereas individual HTTP request executors / HTTP request processors are expected to take care of application specific content generation and processing.

2.8.3.1. Buffering protocol handlers

BufferingHttpServiceHandler and BufferingHttpClientHandler are protocol handler implementations that provide compatibility with the blocking I/O by storing the full content of HTTP messages in memory. Request / response processing callbacks fire only when the entire message content has been read into a in-memory buffer. Please note that request execution / request processing take place the main I/O thread and therefore individual HTTP request executors / request handlers must ensure they do not block indefinitely.

Buffering protocol handler should be used only when dealing with HTTP messages that are known to be limited in length.

2.8.3.2. Throttling protocol handlers

ThrottlingHttpServiceHandler and ThrottlingHttpClientHandler are protocol handler implementations that provide compatibility with the blocking I/O model by utilizing shared content buffers and a fairly small pool of worker threads. The throttling protocol handlers allocate input / output buffers of a constant length upon initialization and control the rate of I/O events in order to ensure those content buffers does not ever overflow. This helps ensure nearly constant memory footprint for HTTP connections and avoid out of memory conditions while streaming content in and out. Request / response processing callbacks fire immediately when a message is received. The throttling protocol handlers delegate the task of processing requests and generating response content to an !Executor, which is expected to perform those tasks using dedicated worker threads in order to avoid blocking the I/O thread.

Usually throttling protocol handlers need only a modest number of worker threads, much fewer than the number of concurrent connections. If the length of the message is smaller or about the size of the shared content buffer worker thread will just store content in the buffer and terminate almost immediately without blocking. The I/O dispatch thread in its turn will take care of sending out the buffered content asynchronously. The worker thread will have to block only when processing large messages and the shared buffer fills up. It is generally advisable to allocate shared buffers of a size of an average content body for optimal performance.

2.8.4. Connection event listener

Protocol handlers like the rest of HttpCore classes do not do logging in order to not impose a choice of a logging framework onto the users. However one can add logging of the most important connection events by injecting a EventListener implementation into the protocol handler.

Connection events as defined by the EventListener interface:

  • fatalIOException Triggered when an I/O error caused the connection to be terminated.

  • fatalProtocolException Triggered when an HTTP protocol error caused the connection to be terminated.

  • connectionOpen Triggered when a new connection has been established.

  • connectionClosed Triggered when the connection has been terminated.

  • connectionTimeout Triggered when the connection has timed out.

2.9. Non-blocking TLS/SSL

2.9.1. SSL I/O session

SSLIOSession is a decorator class intended to transparently extend any arbitrary IOSession with transport layer security capabilities based on the SSL/TLS protocol. Individual protocol handlers should be able to work with SSL sessions without special preconditions or modifications. However, I/O dispatchers need to take some additional actions to ensure correct functioning of the transport layer encryption.

  • When the underlying I/O session has been created, the I/O dispatch must call SSLIOSession#bind() method in order to put the SSL session either into a client or a server mode.

  • When the underlying I/O session is input ready, the I/O dispatcher should check whether the SSL I/O session is ready to produce input data by calling SSLIOSession#isAppInputReady(), pass control to the protocol handler if it is, and finally call SSLIOSession#inboundTransport() method in order to do the necessary SSL handshaking and decrypt input data.

  • When the underlying I/O session is output ready, the I/O dispatcher should check whether the SSL I/O session is ready to accept output data by calling SSLIOSession#isAppOutputReady(), pass control to the protocol handler if it is, and finally call SSLIOSession#outboundTransport() method in order to do the nessary SSL handshaking and encrypt application data.

2.9.1.1. SSL I/O session handler

Applications can customize various aspects of the TLS/SSl protocol by passing a custom implementation of the SSLIOSessionHandler interface.

SSL events as defined by the SSLIOSessionHandler interface:

  • initalize Triggered when the SSL connection is being initialized. The handler can use this callback to customize properties of the javax.net.ssl.SSLEngine used to establish the SSL session.

  • verify Triggered when the SSL connection has been established and initial SSL handshake has been successfully completed. The handler can use this callback to verify properties of the !SSLSession. For instance this would be the right place to enforce SSL cipher strength, validate certificate chain and do hostname checks.

// Get hold of new I/O session
IOSession iosession; 

// Initialize default SSL context
SSLContext sslcontext = SSLContext.getInstance("SSL");
sslcontext.init(null, null, null);

SSLIOSession sslsession = new SSLIOSession(
    iosession, sslcontext, new SSLIOSessionHandler() {

    public void initalize(
        SSLEngine sslengine, 
        HttpParams params) throws SSLException {
        // Ask clients to authenticate
        sslengine.setWantClientAuth(true);
        // Enforce strong ciphers 
        sslengine.setEnabledCipherSuites(new String[] {
                "TLS_RSA_WITH_AES_256_CBC_SHA",
                "TLS_DHE_RSA_WITH_AES_256_CBC_SHA",
                "TLS_DHE_DSS_WITH_AES_256_CBC_SHA" });
    }

    public void verify(
        SocketAddress remoteAddress, 
        SSLSession session) throws SSLException {
        X509Certificate[] certs = session.getPeerCertificateChain();
        // Examine peer certificate chain
        for (X509Certificate cert: certs) {
            System.out.println(cert.toString());
        }
    }
    
}); 

2.9.2. SSL I/O event dispatches

HttpCore provides SSLClientIOEventDispatch and SSLServerIOEventDispatch I/O dispatch implementations that can be used to SSL enable connections managed by any arbitrary I/O reactor. The dispatches take all the necessary actions to wrap active I/O sessions with the SSL I/O session decorator and ensure correct handling of the SSL protocol handshaking.