7122142 : (ann) Race condition between isAnnotationPresent and getAnnotations
peter.levart at gmail.com
Fri Jul 5 08:32:09 UTC 2013
On 07/04/2013 07:34 PM, Joel Borggrén-Franck wrote:
>>> Also, can you please explain to me why the update race is safe. I have done the/some research myself but I would like to know which angles you have covered.
>> Well, one thing is that AnnotationType class is now effectively final, meaning that all it's state is referenced using final fields. If a reference to an instance of such class is obtained via data-race, all it's state is observed consistent by the thread that obtained the reference. The other thing is racy caching. If two or more threads are concurrently requesting AnnotationType for the same Class, many of them might construct and use it's own instance and the one published "latest" will finally be the one being cached. Since all such AnnotationType instances are constructed from the same input data, they are equivalent and it doesn't matter which one is used by which thread.
> Actually they aren't, something have been bugging me and I finally figured it out today. The problem is with default valued elements. Previously all threads have been seeing the same instance of default values but now that will only be guaranteed for Classes, Strings and Integers inside the value cache. While I don't think it is in the spec that annotation default value instances should be == for all threads I don't think it is a race we should introduce. Given this I think you should use some approach to produce a winner that every thread will use (like in the other annotations patch). My gut feeling is that CASing in a result will be better under contention that the current lock solution (while also correct and deadlock free) for a net win.
Here's the 4th revision of the patch:
This one introduces CAS-ing of the AnnotationType instance into the
Class.annotationType field so that there's only a single instance ever
returned for a single Class. I also introduced new private static
Class.Atomic nested class that is used to lazily initialize Unsafe
machinery and to provide a safe wrapper for internal j.l.Class use.
Previously this was part of ReflectionData nested class because it was
only used for it's purpose. In anticipation to have a need for more
atomic operations in the future, I moved this code to a separate class.
ReflectionData is now just a record.
More information about the core-libs-dev