FULL PRODUCT VERSION :
I am using a flavor of java 6.0. The source code which contains the bug:
http://www.java2s.com/Open-Source/Java-Document/6.0-JDK-Core/io-nio/java/nio/channels/spi/AbstractInterruptibleChannel.java.htm

ADDITIONAL OS VERSION INFORMATION :
Linux Ubuntu

A DESCRIPTION OF THE PROBLEM :
The problem is that AbstractInterruptibleChannel#end() does not acquire the closeLock to check the state of interrupted attribute which leaves possibility for a race condition between interrupting the thread and finishing an IO operation. This further causes problems in upper code layers which rely on the fact that the class will throw a ClosedByInterruptException when socket is closed by an interrupt to perform appropriate cleanup.

Also, the interrupt handler should be synchronized with the IO performing thread to avoid channel closure when the thread is already finished.

This bug causes quite rare non-standard behavior of the library which makes it very hard to diagnose for the developers who face it. I spent several days chasing this issue and I think it would be good to fix it just to make sure no-one else will have to debug such anomaly.

STEPS TO FOLLOW TO REPRODUCE THE PROBLEM :
Interrupt threads which perform IO on SelectableChannel and expect one of:
1. ClosedByInterruptException being thrown
2. Socket remaining open and possible to use

EXPECTED VERSUS ACTUAL BEHAVIOR :
EXPECTED -
See above
ACTUAL -
Sometimes a third state appears:
* the socket is closed but the caller is not notified about this state change.

REPRODUCIBILITY :
This bug can be reproduced rarely.