Prior to this change, the ShallowEtagHeaderFilter would use a
ResizableByteArrayOutputStream to internally write data and calculate
the ETag. While that implementation is faster than the regular
ByteArrayOutputStream (since it has a better strategy for growing the
internal buffer), a lot of buffer copying/writing still happens.
This change adds a new FastByteArrayOutputStream implementation that
internally uses a LinkedList<Byte[]> to store the content. So when
writing bytes to that OutputStream implementation, new byte[] are
added to the list when the previous ones are full. This saves most
of the instantiating/copying operations.
Note that new methods were added in DigestUtils to allow usage of
Streams instead of byte[], which is more efficient in our case.
Fixes#653
Issue: SPR-12081
The test was verifying that when a DISCONNECT frame is sent to the
broker, there will be no further messages on the clientOutboundChannel.
This is generally true, however in some cases when the broker receives
a DISCONNECT it may close its connection fast enough (before we do) in
which case we send an ERROR message downstream to ensure the WebSocket
side is cleaned up. Either way the downstream should be idempotent
with regards to cleaning up sessions.
This change designates Jetty SockJS integration tests to run as part of
the "performance", but not the main "publication", CI build due to
recurring low-level failures suspected to be Jetty issues, e.g.
"java.io.IOException: Cannot append to finished buffer" or
"java.io.IOException: Out of order Continuation frame encountered".
The tests will still run at once a day with the performance build but
should not fail the main build with false negatives. Also note that
an Undertow variant of the exact same tests, which hasn't been failing,
will continue to run as part of the main build.
Update HttpComponents wrapper to merge local customizations with the
default of the current HttpClient instead of overriding everything.
This is available as from HttpComponents 4.4. that exposes the default
request config from the client via the Configurable interface. If the
client does not implement such interface, the previous behaviour is
applied
Issue: SPR-12583
Fix a regression introduced by 961574bd17 that prevents a proper lookup
of the default transaction manager by name as the absence of a qualifier
is represented by an empty string (passing the faulty null check).
Issue: SPR-12577
Prior to this commit, some requests would be wrongly mapped to "/" when:
* requests ends with "//" such as "/foo/bar//"
* the DispatcherServlet is mapped as the default ("/")
* the app container sanitizes the servletPath
In those cases, the path extraction algorithm was wrongly guessing the
path of the current request.
This commit detects if the app container sanitized the servletPath for
the current request. If so, it uses a sanitized version of the
requestUri to extract the path information.
Issue: SPR-12372
The following two refinements have been added:
1) SockJS doesn't support binary messages so don't even try
2) don't bother if payload.length == 0
Issue: SPR-12475
Previously the default settings of a custom HttpClient were always
ignored since a RequestConfig instance was always set no matter if
some customizations were applied or not.
This commit keeps an internal RequestConfig object instance that is
only initialized if the user applies a customization. If he does not, the
default settings of the HttpClient are used as it should.
Note that if the HttpComponents API exposed the default RequestConfig
of a given HttpClient, we would be able to merge our customizations with
the one specified by the client. Unfortunately, such API does not exist
and the "defaultSettingsOfHttpClientLostOnExecutorCustomization" test
illustrates that limitation.
Issue: SPR-12540
One more (and hopefully last) attempt at making sure
determineTransactionManager does not break existing use cases.
This commit prevents any lookup if no transaction attributes are set
which is more compliant with the original version and prevents a lookup
if a non existing bean name is provided explicitly (as it can be the case
with Spring Boot).
Issue: SPR-12541
The fix in cec26e9 for SPR-12541 actually introduced a regression
when the interceptor is enabled on a method that does not require any
transaction. In such a case we try to locate the default
PlatformTransactionManager instead of just returning what we have (that
is null).
This commit updates the determineTransactionManager condition again to
take that use case into account again.
Issue: SPR-12541
In order to avoid breaking changed due to the upcoming SockJS 1.0
release (see SPR-12422 for more details) and link to the right
SockJS documentation version, we now explicitly specify that we
support SockJS protocol 0.3.3 and SockJS client 0.3.x.
Before this change attributes listed with @SessionAttributes would not
be saved in the session when there was a redirect and the controller
method declared a parameter of type RedirectAttributes.
This change ensures it's the "default" model that is always the one
checked for @SessionAttributes under all circumstances since
RedirectAttributes is really only meant to provide String values to
insert into or append to the the redirect URL.
Issue: SPR-12542
Previously, if a bean has a scoped proxy and is annotated to be exposed
to the JMX domain, both the scoped proxy and the target instance were
exposed in the JMX domain, resulting in a duplicate entries. Worse, if
such bean defines an explicit name, the application wouldn't start
because of a name conflict.
This commit deals explicitely with scoped proxy and make sure to only
expose the relevant bean.
Issue: SPR-12529
Craig Andrews
Sam Brannen
Stephane Nicoll
John
Rossen Stoyanchev
Juergen Hoeller
Xiaolong Zuo
Brian Clozel
Spring Buildmaster
Sebastien Deleuze