RFR 9: 8138696 : java.lang.ref.Cleaner - an easy to use alternative to finalization
Roger.Riggs at Oracle.com
Fri Oct 2 15:34:10 UTC 2015
Great comments, thanks for looking into this...
On 10/2/2015 7:52 AM, Peter Levart wrote:
> Hi Roger,
> This is a nice and clean API. I like the trick with embedding a
> private CleanerImpl in Cleaner and registering Cleaner itself as a
> tracked object so that automatic thread termination can be performed
> in a safe manner. You say the scope of this is not to replace internal
> usages of sun.misc.Cleaner. But if they ever get replaced, there are
> two differences between the implementations to note:
> - sun.misc.Cleaner thunks get executed by a ReferenceHandler thread
> directly and bypass ReferenceQueue-ing. This might be OK for internal
> use, but not appropriate for public API. I doubt this sun.misc.Cleaner
> exception in ReferenceHandler is necessary though.
> - With sun.misc.Cleaner one has to create a minimum of 2 additional
> objects per tracked object: the Cleaner object and a Runnable thunk.
> This API requires creation of 3 objects per tracked object: the
> Cleanup object, the internal Reference object and a Runnable thunk.
> Do you think Cleaner will not be used in scenarios where this
> additional footprint matters?
I don't have any particular data on that point. When used with lambda
or method references for
the thunk, it is likely there will be some binding overhead.
> It might be possible to merge the roles of Cleanup and Reference into
> one object, like this:
> Making Cleanup an interface, internal Reference(s) can implement it.
> One can cast a Cleanup instance to a Reference and invoke it's other
> methods, but this can be prevented by throwing
> UnsupportedOperationException from them, so nobody is tempted to
> employ this implementation detail.
I prototyped a similar implementation but backed it out due to the code
duplication and complexity.
It also seemed a poor choice to break the contract of Reference by
on the unneeded but exposed methods since they could not be completely
It seemed cleaner to have only a concrete type that was exposed to the
BTW, the code in you cleaner might be simpler if the cleaner lists for
each of the three types were separate.
(The entries don't need to be intermixed). That could simplify the
virtual next/prev access but would still be
duplicating the linked list management code).
There could be multiple insertXXX methods so the casts were not necessary.
> There might be utility in exposing Cleanup (or better named Cleanable)
> References as public abstract classes so that a footprint sensitive
> application (for example in a data structure with many elements) is
> possible. By subclassing such an abstract reference and implementing
> the abstract method, a single object is required per tracked object.
> Here I derived a proof of concept from your code:
> The simple API is unchanged and there is additional low-level API that
> can be used if footprint matters.
> What do you think?
I see what you're getting at. I would probably try to rearrange it so
that only the performMethod
could be overridden. Overriding clean() and clear() might disable the
and reduce the robustness.
I'll take another look, I was most concerned with simplicity, but if the
overhead for the number
of the objects is a concern that can be traded off against a more
> Regards, Peter
> On 10/01/2015 04:12 PM, Roger Riggs wrote:
>> Please review a proposal for public Cleaner API:
>> A Cleaner is proposed to provide an easy to use alternative to
>> finalization. The service would provide easy registration and
>> cancellation of cleanup functions for objects. Applications create a
>> cleanup service for their own use and the service terminates when it
>> is no longer in use.
>> Finalization has a long history of issues both in usage and
>> performance. PhantomReferences have been proposed as the alternative
>> GC based mechanism for cleaning functions but it has been left as an
>> exercise to the developer to construct the necessary mechanisms to
>> handle ReferenceQueues, handle threading issues and robust termination.
>> The Cleaner performs cleaning functions when objects are unreachable
>> as found by garbage collection using the existing mechanisms of
>> PhantomReference, WeakReference, SoftReferences, and ReferenceQueues.
>> It manages a thread that dequeues references to unreachable objects
>> and invokes the corresponding cleaning function. Registered cleaning
>> functions can be cleared if no longer needed, can be invoked
>> explicitly to perform the cleanup immediately, or be invoked when the
>> object is not reachable (as detected by garbage collection) and
>> handled by a cleanup thread.
>> The java.lang.ref package is proposed for the Cleaner because it is
>> complementary to the reference classes and reference queues and to
>> make it easy to find.
>> It is not a goal to replace all uses of finalization or
>> sun.misc.Cleaner in the JDK.
>> Investigation will evaluate if and in what cases the Cleaner can
>> replace finalization.
>> A subsequent task will examine uses of finalization and propose
>> specific changes
>> on a case by base basis.
>> Please review and comment:
>> Thanks, Roger
More information about the core-libs-dev