RFR: 8291023: FileOutputStream.close() re-entrantly invokes itself after getChannel().force()

[Alan Bateman]

On Thu, 6 Apr 2023 22:36:33 GMT, Archie L. Cobbs <duke at openjdk.org> wrote:

> IO stream classes like `FileOutputStream` can have assocated NIO channels.
> 
> When `close()` is invoked on one of these classes, it in turn invokes `close()` on the associated channel (if any). But when the associated channel's `close()` method is invoked, it in turn invokes `close()` on the associated stream (if any).
> 
> As a result, these IO stream `close()` methods invoke themselves reentrantly when there is an associated channel.
> 
> This is not a problem for these classes because they are written to handle this (i.e., they are idempotent), but it can be surprising (or worse, just silently bug-inducing) for subclasses that override `close()`.
> 
> There are two possible ways to address this:
> 1. Modify the code to detect and avoid the (unnecessary) reentrant invocations
> 2. Add a `@implNote` to the Javadoc so subclass implementers are made aware
> 
> This patch takes the second, more conservative approach.

Closing a FIS/FOS/RAF with an associated file channel is specified to close the channel. FileChannel's class description is missing a sentence to say that closing a file channel obtained from a FIS/FOS/RAF will close the stream, something like:


--- a/src/java.base/share/classes/java/nio/channels/FileChannel.java
+++ b/src/java.base/share/classes/java/nio/channels/FileChannel.java
@@ -122,7 +122,8 @@ import jdk.internal.javac.PreviewFeature;
  * the originating object, and vice versa. Changing the file's length via the
  * file channel will change the length seen via the originating object, and vice
  * versa.  Changing the file's content by writing bytes will change the content
- * seen by the originating object, and vice versa.
+ * seen by the originating object, and vice versa.  Closing the channel will
+ * close the originating object.


No objection to adding a note to FIS/FOS/RAF for cases where these classes are extended and close is overridden but I think it should be an API note. As regards JBS issues, the title on JDK-8291023 is misleading as it's nothing to do with FileChannel.close, maybe a new issue should be created for the the API note or else JDK-8291023 it renamed to reflective the re-purpose.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/13379#issuecomment-1500021465