In SPR-16892, the `EncoderHttpMessageWriter` has been improved to write
`"Content-Length"` HTTP response headers if the response body is of type
`Mono` (i.e. the actual content length is easily accessible without
buffering a possibly large response body). That change was relying on
the fact that the server side is using a `ChannelSendOperator` to delay
the writing of the body until the first signal is received.
This strategy is not effective on the client side, since no such channel
operator is used for `WebClient`. This commit improves
`EncoderHttpMessageWriter` and delays, for `Mono` HTTP message bodies
only, the writing of the body so that we can write the
`"Content-Length"` header information once we've got the body resolved.
Issue: SPR-16949
(Cherry-picked from 4a26f93a0d)
This commit fixes the write checks for
`ResourceRegionHttpMessageConverter`, which was previously not checking
properly the parameterized type (e.g. in case of a `List<Something>`).
Issue: SPR-16932
(Cherry-picked from 05ff8b722d)
This commit restricts the allowed HTTP methods on HiddenHttpMethodFilter
(Reactive variant) to the following: PUT, DELETE, PATCH.
This filter is meant to be used to simulate those methods from HTML
forms sent by browsers, so no other methods are allowed.
Issue: SPR-16836
(Cherry-picked from a5cd01a4c8)
This commit restricts the allowed HTTP methods on HiddenHttpMethodFilter
(Servlet variant) to the following: PUT, DELETE, PATCH.
This filter is meant to be used to simulate those methods from HTML
forms sent by browsers, so no other methods are allowed.
Issue: SPR-16836
(Cherry-picked from f64fa3dea1)
Prior to this commit, the generated POMs for Spring Framework modules
would contain unneeded/harmful information from the Spring Framework
build:
1. The BOM imports applied to each module by the dependency
management plugin, for example for Netty or Reactor Netty.
Spring should not export that opinion to its POMs.
2. The exclusion of "org.slf4:jcl-over-slf4j" from *all* dependencies,
which made the POMs much larger than necessary and suggested to
developers that they should exclude it as well when using all those
listed dependencies. In fact, only Apache Tiles currently brings that
transitively.
This commit removes that information from the POMs.
The dependencyManagement Gradle plugin is disabled for POM generation
and we manually resolve the dependency versions during the generation
phase.
The Gradle build is streamlined to exclude "org.slf4:jcl-over-slf4j"
only when necessary.
Issue: SPR-16893
(Cherry-picked from 417354da8a)
With this commit, WebFlux server uses warning instead of error log level
for request handling, and also just print the message instead of the
stacktrace which is mostly meaningless in reactive world.
Complementary to this change, Reactor Netty removed additional logging
as part of https://github.com/reactor/reactor-netty/issues/339.
Issue: SPR-16688
ServerSentEventHttpMessageReader had logic to split on new lines
and buffer until an empty new line (start of a new event). To account
for random data chunking, it later re-assembled the lines for each
event and split again on new lines. However bufferUntil was still
unreliable a chunk may contain nothing but a newline, which doesn't
necessarily mean an empty newline in the overall SSE stream.
This commit simplifies the above by delegating the splitting of the
stream along newlines to StringDecoder.
Issue: SPR-16744
Using the simple example shown in the ticket but switching from
Mono<String> to Flux<String> (and 5,000,000 onNext calls) shows that
constant pausing causes significant overhead and is not worth the
trouble vs ignoring the onWritePossible in REQUESTED state.
Issue: SPR-16702
Undertow does not provide a way to check if we can write so with the
current implementation of isWritePossible, deep recursion can occur
when writing slows down. We now use a flag to keep track of write
ChannelListener callbacks.
This commit also addresses a related issue in
AbstractListenerWriteProcessor that went undected since #3c2d186
where after a large (single) buffer that is not written fully, the
completion signal is processed before the all data is written.
Issue: SPR-16702
This breaks the package dependency cycle between web.server/web.method and makes ServerErrorException more generally applicable. Includes deprecation of the plain reason constructor variant, in favor of providing a Method or MethodParameter context (which MatrixVariableMethodArgumentResolver does now).
Juergen Hoeller
Brian Clozel
Rossen Stoyanchev
Sebastien Deleuze
Johnny Lim
Stephane Maldini
igor-suhorukov