If the response is set and we can't change the status through
ServerHttpResponse any more, allow the error signal to propagate and
let the individual server adapters handle it. Ultimately that should
result in closing the connection.
On Servlet containers, we check one last time if the response is
committed (we may not have filled the buffer). If not then save
the exception as a request attribute, dispatch, and re-throw it on the
container thread.
On Undertow access the connection and close it.
On Netty just let the error through to Reactor Netty.
Issue: SPR-16051
Provide a fully mutable HttpHeaders to ClientHttpRequestInterceptors
of a RestTemplate when headers are set using HttpEntity. This avoids
UnsupportedOperationException if both HttpEntity and
ClientHttpRequestInterceptor add values for the same HTTP header.
Issue: SPR-15066
The request URI returned from HttpServerRequest.uri() typically
contains contains an absolute path but could also contain an
absolute URI. This commit adds handling for the latter, effectively
taking only the absolute path portion.
Issue: SPR-16243
ResponseStatusExceptionHandler lets the error through if it can't
change the status while HttpWebHandlerAdapter logs a more helpful
message (including status code) but without a full stack trace.
Issue: SPR-16231
1. setComplete checks the isCommitted flag to avoid an unnecessary
debug message. This method is meant to be safe to call many times.
2. setStatusCode lowers log message to TRACE, since the return value
communicates the outcome it's arguably much less critical.
3. Add comment and test case for ResponseStatusExceptionHandler.
A ResponseStatusException is clearly meant to be handled by this
handler so don't let it pass through even if the respones is
committed.
Issue: SPR-16231
Before this commit, the return value from write was interpreted as the
data being fully written and ready to be released via releaseData().
This is not true for WebSocketSession implementations where a true
return value simply means the message was sent with the full payload
but releas is not appropriate until a send confirmation.
Technically not an issue since WebSocketSession's extending this do
not use pooled buffers. Nevertheless this commit refines the semantics
of write, removes the releaseData() method, and makes sub-classes
responsible for releasing the buffer when fully written (and they
know best when that is). As a bonus currentData is now private.
Issue: SPR-16207
This commit turns suspendReading() into a readingPaused() notification
that is invoked after a succession of reads stops because there is no
more demand. Sub-classes can use this notification to suspend, if that
applies to them.
Most importantly the notification is guaranteed not to overlap with
checkOnDataAvailable() which means that suspend does not need to be
atomic and guarded against resume. The two can and do compete all the
time when reading ends with no demand, and a request for demand arrives
concurrently.
Issue: SPR-16207
Access-Control-Allow-Credentials CORS header, used to
allow cookies with CORS requests, is not set to true
anymore by default when enabling CORS with
@CrossOrigin or global CORS configuration in order to
provide a more secured default CORS configuration.
The related allowCredentials property now requires to
be set to true explicitly in order to support cookies
with CORS requests.
Issue: SPR-16130
Transition from DEMAND->NO_DEMAND:
Two concurrent threads enter DEMAND.request and DEMAND.onDataAvailable.
And DEMAND.onDataAvailable finishes before DEMAND.request to be able to
update the demand field then a request for reading will be lost.
Transition from READING->NO_DEMAND:
readAndPublish() returns false because there is no demand but before
switching the states READING.request is invoked again a request for
reading will be lost.
Changing READING->DEMAND/NO_DEMAND is made conditional so that the
operations will be executed only if changing states succeeds.
When in READING state detect completion before each next item in order
to exit sooner, if completed.
Issue: SPR-16207
- Ensure completion signal (normal/exception) will be delivered to
the subscriber when transition from UNSUBSCRIBED->COMPLETED
- According to the specification "Publisher.subscribe MUST call onSubscribe
on the provided Subscriber prior to any other signals to that Subscriber" so
ensure onComplete/onError signals will be called AFTER onSubscribe signal.
Issue: SPR-16207
This commit changes the implementation of the PathPattern
extractPathWithinPattern method that used an old AntPathMatcher
derivative to a new version that integrates more closely with
PathContainer.
It also introduces consistency in a couple of areas. The javadoc
is updated to specify this but basically:
- the response from the extra method will have all leading
and trailing separators removed.
- the response will have multiple adjacent separators within the
reponse reduced to just one. (For example response would be
aaa/bb/cc and not aaa///bbb//cc)
If your response would start or finish with multiple separators, they
are all removed.
Issue: SPR-16120
- When there is no demand for reading, the implementation
should suspend reading otherwise useless events will be
send by Undertow to the registered read listener.
- There is not need to wait for an event for reading/writing after calling
resumeReads/resumeWrites
Practically no changes to RxNetty for a year and efforts underway to
rebuild 0.6.x based on a current Reactor Netty base.
Aside from the extra time to run integration tests having two
Netty-based servers can also cause false alarms such as ByteBuf leaks
related to RxNetty.
Also remove ErrorDeferredResultProcessingInterceptor which was meant
to be package private to begin with as its peer
ErrorCallableProcessingInterceptor (also removed).
Rossen Stoyanchev
Arjen Poutsma
Juergen Hoeller
sdeleuze
Christoph Dreis
Sam Brannen
Violeta Georgieva
Andy Clement