Prior to this commit, the ServletResponseHttpHeaders.get method
would throw an NPE when used under Wildfly 8.0.0.Final and 8.1.0.Final.
This can be traced to WFLY-3474, which throws an NPE when calling
HttpServletResponse.getHeaders("foo") and that header has not
been defined prior to that.
This would cause NPE being thrown by AbstractSockJsService when
checking for CORS HTTP headers in the server HTTP response.
This commit surrounds that method call in AbstractSockJsService and
guards against this issue.
Issue: SPR-11919
In a recent CI build failure a test timed out waiting for a message:
https://build.spring.io/browse/SPR-PUB-1316
In fact there was a WebSocket transport failure originating in Jetty's
WebSocket Parser. However this failure is only apparent by looking at
the logs. This change adds a check if a transport error ocurred while
we were waiting and throws an AssertionError.
Sessions connected to a STOMP endpoint are expected to receive some
client messages. Having received none after successfully connecting
could be an indication of proxy or network issue. This change adds
periodic checks to see if we have not received any messages on a
session which is an indication the session isn't going anywhere
most likely due to a proxy issue (or unreliable network) and close
those sessions.
Issue: SPR-11884
Commit 5d2e6f enabled the setRemoveOnCancelPolicy of the SockJS
ScheduledThreadPoolExecutor by default. However that property is only
available in JDK 1.7 or later. This change fixes the issue and
introduces an extension of ThreadPoolTaskScheduler for for use with
SockJS which is necessary in any case since ThreadPoolTaskScheduler
does not expose the underlying ScheduledThreadPoolExecutor otherwise.
Issue: SPR-11918
This change sets the removeOnCancelPolicy on the SockJS
ScheduledThreadPoolExecutor to true. This ensures that cancelled
tasks are removed immediately to avoid the "unbounded retention
of cancelled tasks" that is mentioned in the Javadoc of
ScheduledThreadPoolExecutor:
"By default, such a cancelled task is not automatically removed from
the work queue until its delay elapses. While this enables further
inspection and monitoring, it may also cause unbounded retention of
cancelled tasks. To avoid this, set setRemoveOnCancelPolicy to true,
which causes tasks to be immediately removed from the work queue at
time of cancellation."
Issue: SPR-11918
This change ensures the server "WebSocketHandler" is notified of the
opening of a session before writing the open frame to the remote
handler. Any messages sent by the server "WebSocketHandler" while
getting notified of the opening get cached and flushed after the open
frame has been written.
This change introduces locking in AbtractHttpSockJsSession to guard
access to the HTTP response. The goal is to prevent contention between
client requests to receive messages (i.e. long polling) and
the application trying to write.
Issue: SPR-11916
This change adds a new implementation of WebSocketClient that can
connect to a SockJS server using one of the SockJS transports
"websocket", "xhr_streaming", or "xhr". From a client perspective
there is no implementation difference between "xhr_streaming" and
"xhr". Just keep receiving and when the response is complete,
start over. Other SockJS transports are browser specific
and therefore not relevant in Java ("eventsource", "htmlfile" or
iframe based variations).
The client loosely mimics the behavior of the JavaScript SockJS client.
First it sends an info request to find the server capabilities,
then it tries to connect with each configured transport, falling
back, or forcing a timeout and then falling back, until one of the
configured transports succeeds.
The WebSocketTransport can be configured with any Spring Framework
WebSocketClient implementation (currently JSR-356 or Jetty 9).
The XhrTransport currently has a RestTemplate-based and a Jetty
HttpClient-based implementations. To use those to simulate a large
number of users be sure to configure Jetty's HttpClient executor
and maxConnectionsPerDestination to high numbers. The same is true
for whichever underlying HTTP library is used with the RestTemplate
(e.g. maxConnPerRoute and maxConnTotal in Apache HttpComponents).
Issue: SPR-10797
This change adds support for a custom "websocket" scope.
WebSocket-scoped beans may be injected into controllers with message
handling methods as well as channel interceptor registered on the
"inboundClientChannel".
Issue: SPR-11305
Before this change, subscribing to a user destination and use of
@SendToUser annotation required an authenticated user.
This change makes it possible to subscribe to a user destination from
WebSocket sessions without an authenticated user. In such cases the
destination is associated with one session only rather than with a
user (and all their sessions).
It is then also possible to send a message to a user destination
via "/user/{sessionId}/.." rather than "/user/{user}/...".
That means @SendToUser works relying on the session id of the input
message, effectively sending a reply to destination private to the
session.
A key use case for this is handling an exception with an
@MessageExceptionHandler method and sending a reply with @SendToUser.
Issue: SPR-11309
Animal sniffer provides tools to assist verifying that classes
compiled with a newer JDK are compatible with an older JDK.
This integratesthe latest version of the tool (1.11) that
permits the use of custom annotations. Added @UsesJava7,
@UsesJava8 and @UsesSunHttpServer and annotated the few places
where we rely on a specific environment.
The verification process can be invoked by running the 'sniff'
task.
Issue: SPR-11604
polishing
The original fix for SPR-11423:
32e5f57e64
was insufficient when using an external broker since the original
destination header has to be in the "native headers" map (i.e. with
STOMP headers) in order to be included in messages broadcast by
the broker.
The BufferingStompDecoder now decorates rather than extend
StompDecoder. This allows a single StompDecoder instance to be
configured and extended independantly while buffering remains a
separate concern.
Mutate rather than re-create headers when decoding STOMP messages
before a message is sent on a message channel.
Use MessageBuilder.createMessage to ensure the fully prepared
MessageHeaders is used directly MessageHeaderAccessor instance.
Issue: SPR-11468
Refine semantics of ID and TIMESTAMP headers provided to protected
MessageHeaders constructor.
Refactor internal implementation of MessageHeaderAccessor.
Support mutating headers from a single thread while a message is being
built (e.g. StompDecoder creating message + then adding session id).
Improve immutablity in NativeMessageHeaderAccessor and in
StompHeaderAccessor.
Optimize object creation for initializing messages and subsequent
accessing their headers.
Introduce MessageHeaderAccessorFactory support to enable applying a
common strategies for ID and TIMESTAMP generation to every message.
Add MessageBuilder shortcut factory method for creating messages from
payload and a full-prepared MessageHeaders instance. Also add
equivalent constructors to GenericMessage and ErrorMessage.
Issue: SPR-11468
The UserDestinationMessageHandler adds a header providing a hint for
what the original destination a user may have used when subscribing.
That is then used when writing messages back to WebSocket clients to
ensure they dont see the internally used, transformed user destination.
This change moves the header name constatn to make it more broadly
applicable. For example SPR-11645.
Prior to this commit, @SubscribeMapping mapped methods (backed with
@SendTo* annotations, or not) would send MESSAGEs with the wrong
destination. Instead of using the original SUBSCRIBE destination, it
would use the lookup path computed from the configured prefixes in the
application.
This commit fixes this issue - now @SubscribeMapping MESSAGEs use the
original SUBSCRIBE destination.
Issue: SPR-11648
Update some native WebSocket session getters to return basic
information after it is closed. It is required for example in
SubProtocolWebSocketHandler#afterConnectionEstablished() or
StompSubProtocolHandler#afterSessionStarted().
Issue: SPR-11621
Proactively notify all active WebSocket sessions when a shutdown is
progress. Sessions then can ignore further attempts to send messages
and also stop stop trying to flush messages right away.
BufferingStompDecoder message buffer size limit can now be configured
with JavaConfig MessageBrokerRegistry.setMessageBufferSizeLimit() or
with XML <websocket:message-brocker message-buffer-size="">.
Issue: SPR-11527
Before this change the StompDecoder decoded and returned only the first
Message in the ByteBuffer passed to it. So to obtain all messages from
the buffer, one had to loop passing the same buffer in until no more
complete STOMP frames could be decoded.
This chage modifies StompDecoder to return List<Message> after
exhaustively decoding all available STOMP frames from the input buffer.
Also an overloaded decode method allows passing in Map that will be
populated with any headers successfully parsed, which is useful for
"peeking" at the "content-length" header.
This change also adds a BufferingStompDecoder sub-class which buffers
any content left in the input buffer after parsing one or more STOMP
frames. This sub-class can also deal with fragmented messages,
re-assembling them and parsing as a whole message.
Issue: SPR-11527
When a send timeout is detected, the WebSocket session is now closed
with a custom close status that indicates so. This allows skipping
parts of the close logic that may cause further hanging.
Issue: SPR-11450
Before this change SockJsSession implementations of WebSocketSession
used synchronization around its method implementations protecting
internal state and ensuring only a single thread is sending messages
at a time.
A WebSocketSession is generally expected to be used from one thread
at a time and now that application messages are sent through
ConcurrentWebSocketSessionDecorator, there is no concern about
application messages sent from the different threads.
While there are some remaining concerns, those can be addressed
without using the synchronized keyword. This change removes it from
the methods of all SockJS session implementations.
Issue: SPR-11450
Before this change the decorator ensured that for a specific WebSocket
session only one thread at a time can send a message. Other threads
attempting to send would have their messages buffered and each time
that occurs, a check is also made to see if the buffer limit has been
reached or the send time limit has been exceeded and if so the session
is closed.
This change adds further protection to ensure only one thread at a time
can perform the session limit checks and attempt to close the session.
Furthermore if the session has timed out and become unresponsive,
attempts to close it may block yet another thread. Taking this into
consideration this change also ensures that state associated with the
session is cleaned first before an attempt is made to close the session.
Issue: SPR-11450
Since we now wrap the WebSocketSession with a concurrent decorator, the
synchronized keyword around message sending needed to be removed.
Issue: SPR-11586
Prior to this commit, configuring a custom handshakeHandler when setting
up a stomp-endpoint with SockJS would not be taken into account:
<websocket:stomp-endpoint path="/foo">
<websocket:handshake-handler ref="customHandler"/>
<websocket:sockjs/>
</websocket:stomp-endpoint>
This commit fixes this by creating and registering a
WebsocketTransportHandler (with this handshakeHandler) as a
transportHandler override for the SockJSService.
Issue: SPR-11568
Prior to this commit, the codebase was using a mix of log4j.xml
and log4j.properties for test-related logging configuration. This
can be an issue as log4j takes the xml variant first when looking
for a default bootstrap configuration.
In practice, some modules declaring the properties variant were
taking the xml variant configuration from another module.
The general structure of the configuration has also been
harmonized to provide a standard console output as well as an
easy way to enable trace logs for the current module.
The clientInboundChannel and clientOutboundChannel now use twice
the number of available processors by default to accomodate for some
degree of blocking in task execution on average.
In practice these settings still need to be configured explicitly in
applications but these should serve as better default values than
the default values in ThreadPoolTaskExecutor.
Issue: SPR-11450
This change exposes the WebSocketSession attributes through a message header.
The StompSubProtocolHandler adds this to incoming messages.
For now messaging handling methods can access the map via @Header, e.g.:
@Header(StompHeaderAccessor.SESSION_ATTRIBUTES) Map<String, Object> attrs) {
Issue: SPR-11566
Add accessor for brokerAvailable in AbstractBrokerMessageHandler
Ensure brokerAvailable is set even if eventPublisher is not
Add tests BrokerMessageHandlerTests
Turn off brokerAvailable when StompBrokerRelayMessageHandler stops
Actually stop message handling when brokerAvailable=false
Improve log messages
Issue: SPR-11563
Prior to this commit, the `relay-port` attribute of the
`<websocket:stomp-broker-relay />` tag was of type `xsd:int`.
This prevents developers from using `PropertyPlaceholderConfigurer`,
even though this configuration key is a good candidate for such use
(this value depends on prod/staging/etc environment).
This commit changes that type to `xsd:string`.
Issue: SPR-11537
After this change, AbstractSockJsService does not add CORS headers if
the response already contains an "Access-Control-Allow-Origin" header.
Essentially it backs off assuming CORS headers are handled centrally
e.g. through a Filter.
In order to support this, the ServletServerHttpResponse now returns an
instance of HttpHeaders that also provides access to headers already
present in the HttpServletResponse.
Issue: SPR-11443
- Deleted empty AbstractWebSocketClientTests class.
- AbstractServletHandlerMethodTests and AbstractHttpRequestTests are
now actually declared as abstract.
- The following classes are not abstract but currently have an
"Abstract" prefix and therefore get ignored by the Gradle build.
This commit renames each of these by deleting the "Abstract" prefix.
- AbstractFlashMapManagerTests
- AbstractMappingContentNegotiationStrategyTests
- AbstractSockJsServiceTests
- AbstractWebSocketHandlerRegistrationTests
The Servlet API does not provide notifications when a client
disconnects, see see https://java.net/jira/browse/SERVLET_SPEC-44.
Therefore network IO failures may occur simply because a client has
gone away. Before this change that could fill logs with unnecessary
stack traces.
After this change we make a best effort to identify such network
failures, on a per-server basis (tested with Jetty, Tomcat, Glassfish,
and WildFly), and log them under a separate log category.
A simple one-line message is logged at DEBUG level (i.e. no stack trace)
while a full stack trace is shown at TRACE level.
Issue: SPR-11438
Before this change CompositeMessageConverter had a ContentTypeResolver
field that was in turn set on all contained converters.
After this change that field is removed and effectively
CompositeMessageConverter is a simple container of other converters.
Each converter in turn must have been configured with a
ContentTypeResolver.
Doing so means it is less likely to have unexpected consequences when
configuring converters, the ContentTypeResolver set in the composite
converter overriding the one configured in a contained converter.
Also commit 676ce6 added default ContentTypeResolver initialization
to AbstractMessageConverter, which ensures that converters are still
straight forward to configure.
Issue: SPR-11462
Before this change, when a client subscribed to a "user" destination
(e.g. /user/foo), actual messages received in response to that
subscription contained the server-translated, unique user destination
(e.g. /foo-user123).
This is not an issue for clients such as stomp.js since the
subscription is unique and sufficient to match subscription responses.
However, other STOMP clients do additional checks on the destination
of the subscription and the response.
This change ensures that messages sent to clients on user destionations
always contain a destination that matches the one on the original
subscription.
Issue: SPR-11423
The interface is to be implemented in addition to
java.security.Principal when Principal.getName() is not globally unique
enough for use in user destinations.
Issue: SPR-11327
sockjs-client expects a prelude to be written on every request with
streaming transports. The protocol tests don't make this clear and
don't expose this issue.
The test case for SPR-11183 (writing 20K messages in succession) did
expose the issue and this commit addresses it.
Issue: SPR-11183
Brian Clozel
Rossen Stoyanchev
Juergen Hoeller
Artem Bilan
Stephane Nicoll
Philippe Marschall
Sebastien Deleuze
Sam Brannen