RFR 9: 8138696 : java.lang.ref.Cleaner - an easy to use alternative to finalization
Peter Levart
peter.levart at gmail.com
Tue Oct 6 16:06:54 UTC 2015
Hi Roger (again, sorry)!
I think I found a race in the implementation which could leave the
Cleaner thread run forever:
238 public void run() {
239 Thread t = Thread.currentThread();
240 ManagedLocalsThread mlThread = (t instanceof
ManagedLocalsThread)
241 ? (ManagedLocalsThread)t
242 : null;
243 while (!phantomCleanableList.isEmpty() ||
244 !weakCleanableList.isEmpty() ||
245 !softCleanableList.isEmpty()) {
246 if (mlThread != null) {
247 // Cleanable the thread locals
248 mlThread.eraseThreadLocals();
249 }
250 try {
251 Cleanable ref = (Cleanable) queue.remove();
252 ref.clean();
253 } catch (InterruptedException i) {
254 continue; // ignore the interruption
255 } catch (Throwable e) {
256 // ignore exceptions from the cleanup thunk
257 }
258 }
259 }
Suppose the Cleaner instance has already been collected and it's
associated PhantomCleanableRef has already been removed from
phantomCleanableList. Let weakCleanableList and softCleanableList be
(already) empty too. Let there be some last PhantomCleanableRefefence in
the phantomCleanableList. while condition is therefore true, so the
Cleaner thread enters loop and blocks at queue.remove(). Now some user
thread clean()s or clear()s the remaining PhantomCleanableRefefence
before it's referent is found phantom reachable by GC and removes the
Reference from the list. This reference will never be enqueued by
GC+ReferenceHandler, so queue.remove() will block forever.
The simplest fix would be to use a remove() with a timeout in this
situation.
Regards, Peter
On 10/06/2015 12:11 AM, Roger Riggs wrote:
> Hi,
>
> Building on Peter's code to provide subclassable CleanableReferences,
> I rearranged
> the implementation a bit to make it easier to maintain.
> The CleanableReferences subclasses are nested inside the Cleaner to
> make the scoping clear.
>
> Please review and comment:
>
> Webrev:
> http://cr.openjdk.java.net/~rriggs/webrev-cleaner-extensible-8138696/
>
> javadoc:
> http://cr.openjdk.java.net/~rriggs/cleaner-doc2/
>
> Thanks, Roger
>
>
>
> On 10/2/2015 11:34 AM, Roger Riggs wrote:
>> Hi Peter,
>>
>> 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:
>>>
>>> http://cr.openjdk.java.net/~plevart/jdk9-dev/Cleaner/Cleaner.java
>>>
>>> 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
>> throwing UnsupportedOperationException
>> on the unneeded but exposed methods since they could not be
>> completely encapsulated.
>> It seemed cleaner to have only a concrete type that was exposed to
>> the clients.
>>
>> 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:
>>>
>>> http://cr.openjdk.java.net/~plevart/jdk9-dev/Cleaner/webrev.01/
>>>
>>> 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 list management
>> 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
>> complex implementation.
>>
>> Thanks, Roger
>>
>>>
>>> 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:
>>>>
>>>> Javadoc:
>>>> http://cr.openjdk.java.net/~rriggs/cleaner-doc/
>>>>
>>>> Webrev:
>>>> http://cr.openjdk.java.net/~rriggs/webrev-cleaner-8138696/
>>>>
>>>> Issue:
>>>> https://bugs.openjdk.java.net/browse/JDK-8138696
>>>>
>>>> Thanks, Roger
>>>>
>>>
>>
>
More information about the core-libs-dev
mailing list