springloaded isn't required in any of the samples, yet some of them
make use of it without being a springloaded-specific sample. This
is creating the false impression that springloaded is necessary in
cases where it's not.
When a custom management.port is used, the child context is configured
with an EmbeddedServletContainerFactory bean that has the same class
as the parent context’s EmbeddedServletContainerFactory bean. This
ensures that the child context uses the same type of embedded container
as its parent when there are multiple embedded containers on the
classpath. It also causes a failure when the custom
EmbeddedServletContainerFactory subclass cannot be instantiated, for
example because it’s an anonymous inner-class.
This commit improves the diagnostics so that we fail fast with an
information exception message when we detect that the embedded servlet
container factory bean’s class cannot be instantiated.
Closes gh-6193
Previously, if the `contextPath` of the application wasn't the root, the
HAL browser could not initialize since the `entryPoint` was referring to
an invalid location.
This commit makes sure to take the `contextPath` into account.
Closes gh-5814
Previously, SimpleInMemoryRepository used a ConcurrentReferenceHashMap
to store its locks. The type of map will discard its entries when the
JVM comes under GC pressure. With the code in its previous form, this
could lead to a NullPointerException when the following occurred:
1. putIfAbsent returned null indicating that a new entry has been added
to the map
2. GC pressure caused the map to discard the new entry
3. get returned null as the entry has been discard
There are two problems with the existing code:
1. Its usage of a ConcurrentMap is incorrect. The correct usage is:
a. Call get to see if the map already contains a lock
b. If the lock is null, create a new one
c. Call putIfAbsent to add the new lock
d. If the return value is non-null, another thread has created the
lock and it should be used. If the return value is null, use the
new lock created in b.
2. Once the use of ConcurrentMap has been corrected, the fact that it is
a ConcurrentReferenceHashMap means that different threads could
access the same value using different locks. This would occur if one
thread has retrieved a lock from the map and is using it, while GC
causes the lock to be removed from the map. Another thread then
attempts to get the lock and, as GC pressure has remove it, a new
lock is created allowing concurrent access to the same value.
This commit updates the code to use the ConcurrentMap correctly and also
replaces the ConcurrentReferenceHashMap with a ConcurrentHashMap. This
means that the repository will now use slightly more memory but this is
outweighed by the benefits of thread-safe updates and no risk of an NPE.
Closes gh-6115
DevTools deliberately throws an uncaught exception on the main thread
as a safe way of causing it to stop processing. This exception is
caught and swallowed by an uncaught exception handler. Unfortunately,
this has the unwanted side-effect of causing the JVM to exit with 1
once all running threads are daemons.
Normally, this isn't a problem. Non-daemon threads, such as those
started by an embedded servlet container, will keep the JVM alive and
restarts of the application context will occur when the user makes to
their application. However, if the user adds DevTools to an
application that doesn't start any non-daemon threads, i.e. it starts,
runs, and then exits, it will exit with 1. This causes both
bootRun in Gradle and spring-boot:run in Maven to report that the
build has failed. While there's no benefit to using DevTools with an
application that behaves in this way, the side-effect of causing the
JVM to exit with 1 is unwanted.
This commit address the problem by updating the uncaught exception
handler to call System.exit(0) if the JVM is going to exit as a
result of the uncaught exception causing the main thread to die. In
other words, if the main thread was the only non-daemon thread, its
death as a result of the uncaught exception will now cause the JVM
to exit with 1 rather than 0. If there are other non-daemon threads
that will keep the JVM alive, the behaviour is unchanged.
Closes gh-5968
In fact the folder was already created if the app is running as
a different user, but not if running as the current user, so it
was just a question of moving one line out of an if block.
Fixes gh-5986
Update SpringApplicationBuilder so that properties of the form
`abc=d:e:f` are correctly parsed. Prior to this commit the `:` delimiter
would always be chosen over `=`, even if `=` occurred first.
Fixes gh-6121
Update TypeUtils to guard against the use of older Java versions.
Both `Collection` and `Map` type lookups now fallback to generic free
versions of the classes.
Prior to this commit using `xmlbeans-maven-plugin` in combination with
Spring Boot's annotation processor could result in
`IllegalArgumentException: Incorrect number of type arguments`.
Fixes gh-6122
Update BeanDefinitionLoader to support loading from package names that
do not contain dots.
Prior to this commit `new BeanDefinitionLoader(registry, "somepackage")`
would fail because "somepackage" exists and is a resource but does not
contain valid XML. Somewhat surprisingly the InputStream returned by
the resource actually contains the listing of files in the package.
Fixes gh-6126
- Limit shared state between FileSystemWatcher and the watching thread
- Use a private monitor rather than synchronizing on this
- Use a Runnable implementation rather than subclassing Thread
- Synchronize consistently when reading and writing state
Closes gh-6039
Previously, JarURLConnection didn't override getPermission(). This
meant that it required all permissions. This was at odds with the
Oracle JVM's concrete sun.net.www.protocol.jar.JarURLConnection which
overrides getPermission to return a FilePermission with the read
action for the path of the underlying jar.
This commit updates our JarURLConnection to align its behaviour with
sun.net.www.protocol.jar.JarURLConnection.
Closes gh-5411
The list of changes is written to on one thread and read from on
another. Without some form of sychronization this is not thread-safe.
This commit makes changes a synchronized list which should guarantee
that the reading thread can see the changes made by the writing thread.
It also removes a redundant call to clear the list of changes at the
start of waitsForPollingInterval.
See gh-6038
Neither Cassandra nor Elasticsearch starts reliably on Windows. This
commit adds a custom class rule to the associated sample application
tests to skip them on Windows. A class rule is used rather than a
Unit assumption as we want to avoid starting Elasticsearch (done by
the application context) and Cassandra (done by a test execution
listener) and an assumption would be too late.