When concurrency limiting is enabled via setConcurrencyLimit() and
thread creation fails in doExecute() (e.g., OutOfMemoryError from
Thread.start()), the concurrency permit acquired by beforeAccess()
is never released because TaskTrackingRunnable.run() never executes.
This causes the concurrency count to permanently remain at the limit,
causing all subsequent task submissions to block forever in
ConcurrencyThrottleSupport.onLimitReached().
Root cause:
- beforeAccess() increments concurrencyCount
- doExecute() throws Error before thread starts
- TaskTrackingRunnable.run() never executes
- afterAccess() in finally block never called
- Concurrency permit permanently leaked
Solution:
Wrap doExecute() in try-catch block in the concurrency throttle path
and call afterAccess() in catch block to ensure permit is always
released, even when thread creation fails.
The fix only applies to the concurrency throttle path. The
activeThreads-only path does not need fixing because it never calls
beforeAccess(), so there is no permit to leak.
Test approach:
The test simulates thread creation failure and verifies that a
subsequent execution does not deadlock. The first execution should
fail with some exception (type doesn't matter), and the second
execution should complete within timeout if the permit was properly
released.
Signed-off-by: Park Juhyeong <wngud5957@naver.com>
Previously, SingletonSupplier stored "null" in singletonInstance when
the supplied instance was "null". On subsequent get() calls, this was
treated as "uninitialized" and triggered another attempt to obtain an
instance from the Supplier.
This commit ensures that a "null" returned from the instanceSupplier or
defaultSupplier is handled correctly, so that subsequent calls to get()
return "null" consistently instead of repeatedly invoking the Supplier.
Closes gh-35380
Signed-off-by: Dmytro Nosan <dimanosan@gmail.com>
Prior to this commit, getPubliclyAccessibleMethodIfPossible() in
ClassUtils incorrectly returned a hidden static method as an
"equivalent" method for a static method with the same signature;
however, a static method cannot be overridden and therefore has no
"equivalent" method in a super type.
To fix that bug, this commit immediately aborts the search for an
"equivalent" publicly accessible method when the original method is a
static method.
See gh-33216
See gh-35189
See gh-35556
Closes gh-35667
As reported in gh-34651, `DataBuffer#getByte` can be inefficient for
some implementations, as bound checks are performed for each call.
This commit introduces a new `forEachByte` method that helps with
traversing operations without paying the bound check cost for each byte.
Closes gh-35623
Prior to this commit, our @Retryable support as well as a RetryPolicy
created by the RetryPolicy.Builder only matched against top-level
exceptions when filtering included/excluded exceptions thrown by a
@Retryable method or Retryable operation.
With this commit, we now match against not only top-level exceptions
but also nested causes within those top-level exceptions. This is
achieved via the new ExceptionTypeFilter.match(Throwable, boolean)
support.
See gh-35592
Closes gh-35583
Prior to this commit, ExceptionTypeFilter only provided support for
filtering based on exact matches against exception types; however, some
use cases require that filtering be applied to nested causes in a given
exception. For example, this functionality is a prerequisite for
gh-35583.
This commit introduces a new match(Throwable, boolean) method in
ExceptionTypeFilter, where the boolean flag enables matching against
nested exceptions.
See gh-35583
Closes gh-35592
SingletonSupplier<T> supplier-based static methods nullability should
be refined to accept in a flexible way nullable or non-nullable T.
Closes gh-35559
Juergen Hoeller