ThreadPoolExecutor and finalization
Roger Riggs
Roger.Riggs at Oracle.com
Tue Oct 31 13:58:05 UTC 2017
Hi Peter,
Only native resources that do not map to the heap allocation/gc cycle
need any kind
of cleanup. I would work toward a model that encapsulates the reference
to a native resource
with a corresponding allocation/release mechanism as you've described
here and in the
thread on zip.
For cleanup purposes, the independence of each resource may improve
robustness
by avoiding dependencies and opportunities for entanglements and bugs
due to exceptions
and other failures.
In the case of TPE, the native resources are Threads, which keep running
even if they are
unreferenced and are kept referenced via ThreadGroups.
I don't know the Executor code well enough to do more than speculate,
but would suggest
that a cleaner (or similar) should be registered for each thread .
For TPE, since Threads do not become unreferenced, the part of the spec
related to finalize
being called by the finalizer thread is moot.
$.02, Roger
On 10/31/2017 5:24 AM, Peter Levart wrote:
> Hi,
>
> Here are some of my thoughts...
>
> On 10/31/17 05:37, David Holmes wrote:
>> Hi Roger,
>>
>> On 31/10/2017 12:43 AM, Roger Riggs wrote:
>>> Hi David,
>>>
>>> On 10/30/2017 3:31 AM, David Holmes wrote:
>>>> Hi Andrej,
>>>>
>>>> On 30/10/2017 5:02 PM, Andrej Golovnin wrote:
>>>>> Hi David,
>>>>>
>>>>>> On 30. Oct 2017, at 01:40, David Holmes <david.holmes at oracle.com>
>>>>>> wrote:
>>>>>>
>>>>>> Hi Roger,
>>>>>>
>>>>>> On 30/10/2017 10:24 AM, Roger Riggs wrote:
>>>>>>> Hi,
>>>>>>> With the deprecation of Object.finalize its time to look at its
>>>>>>> uses too see if they can be removed or mitigated.
>>>>>>
>>>>>> So the nice thing about finalize was that it followed a
>>>>>> nice/clean/simple OO model where a subclass could override, add
>>>>>> their own cleanup and then call super.finalize(). With finalize()
>>>>>> deprecated, and the new mechanism being Cleaners, how do Cleaners
>>>>>> support such usages?
>>>>>
>>>>> Instead of ThreadPoolExecutor.finalize you can override
>>>>> ThreadPoolExecutor.terminated.
>>>>
>>>> True. Though overriding shutdown() would be the semantic equivalent
>>>> of overriding finalize(). :)
>>>>
>>>> In the general case though finalize() might be invoking a final
>>>> method.
>
> Overriding shutdown() would only work when this method is explicitly
> invoked by user code. Using Cleaner API instead of finalization can
> not employ methods on object that is being tracked as that object is
> already gone when Cleaner invokes the cleanup function.
>
> Migrating TPE to Cleaner API might be particularly painful since the
> state needed to perform the shutdown is currently in the TPE instance
> fields and would have to be moved into the cleanup function. The
> easiest way to do that is to:
>
> - rename class ThreadPoolExecutor to package-private
> ThreadPoolExecutorImpl
> - create new class ThreadPoolExecutor with same public API delegating
> the functionality to the embedded implementation
> - registering Cleaner.Cleanable to perform shutdown of embedded
> executor when the public-facing executor becomes phantom reachable
>
> This would have an added benefit of automatically shut(ing)down() the
> pool when it is not reachable any more but still has idle threads.
>
>>>>
>>>> Anyway I'm not sure we can actually do something to try to move
>>>> away from use of finalize() in TPE. finalize() is only deprecated -
>>>> it is still expected to work as it has always done. Existing
>>>> subclasses that override finalize() must continue to work until
>>>> some point where we say finalize() is not only deprecated but
>>>> obsoleted (it no longer does anything). So until then is there
>>>> actually any point in doing anything? Does having a Cleaner and a
>>>> finalize() method make sense? Does it aid in the transition?
>>> As you observe, the alternatives directly using PhantomRefs or the
>>> Cleanup do not provide
>>> as nice a model. Unfortunately, that model has been recognized to
>>> have a number of
>>> issues [1]. Finalization is a poor substitute for explicit
>>> shutdown, it is unpredictable and unreliable,
>>> and not guaranteed to be executed.
>>
>> I'm not trying to support/defend finalize(). But it does support the
>> very common OO pattern of "override to specialize then call super".
>
> I have been thinking about that and I see two approaches to enable
> subclasses to contribute cleanup code:
>
> - one simple approach is for each subclass to use it's own independent
> Cleaner/Cleanable. The benefit is that each subclass is independent of
> its super/sub classes, the drawback is that cleanup functions are
> called in arbitrary order, even in different threads. But for
> cleaning-up independent resources this should work.
>
> - the other approach is more involving as it requires the base class
> to establish an infrastructure for contributing cleanup code. For this
> purpose, the internal low-level Cleaner API is most appropriate
> thought it can be done with public API too. For example (with public API):
>
>
> public class BaseClass implements AutoCloseable {
>
> protected static class BaseResource implements Runnable {
> @Override
> public void run() {
> // cleanup
> }
> }
>
> protected final BaseResource resource = createResource();
> private final Cleaner.Cleanable cleanable =
> CleanerFactory.cleaner().register(this, resource);
>
> protected BaseResource createResource() {
> return new BaseResource();
> }
>
> public BaseClass() {
> // init BaseResource resource...
> }
>
> @Override
> public final void close() {
> cleanable.clean();
> }
> }
>
>
> public class DerivedClass extends BaseClass {
>
> protected static class DerivedResource extends BaseResource {
> @Override
> public void run() {
> // before-super cleanup
> super.run();
> // after-super cleanup
> }
> }
>
> @Override
> protected DerivedResource createResource() {
> return new DerivedResource();
> }
>
> public DerivedClass() {
> super();
> DerivedResource resource = (DerivedResource) this.resource;
> // init DerivedResource resource
> }
> }
>
>
> And alternative with private low-level API:
>
>
> public class BaseClass implements AutoCloseable {
>
> protected static class BaseResource extends
> PhantomCleanable<BaseClass> {
>
> protected BaseResource(BaseClass referent) {
> super(referent, CleanerFactory.cleaner());
> }
>
> @Override
> protected void performCleanup() {
> // cleanup
> }
> }
>
> protected final BaseResource resource = createResource();
>
> protected BaseResource createResource() {
> return new BaseResource(this);
> }
>
> public BaseClass() {
> // init BaseResource resource...
> }
>
> @Override
> public final void close() {
> resource.clean();
> }
> }
>
>
> public class DerivedClass extends BaseClass {
>
> protected static class DerivedResource extends BaseResource {
>
> protected DerivedResource(DerivedClass referent) {
> super(referent);
> }
>
> @Override
> protected void performCleanup() {
> // before-super cleanup
> super.performCleanup();
> // after-super cleanup
> }
> }
>
> @Override
> protected DerivedResource createResource() {
> return new DerivedResource(this);
> }
>
> public DerivedClass() {
> super();
> DerivedResource resource = (DerivedResource) this.resource;
> // init DerivedResource resource
> }
> }
>
>
> Regards, Peter
>
>>
>>> ThreadPoolExecutor has a responsibility to cleanup any native
>>> resources it has allocated (threads)
>>> and it should be free to use whatever mechanism is appropriate.
>>> Currently, the spec for finalize
>>> does not give it that freedom.
>>
>> I suppose in hindsight we (JSR-166 EG) could have described automatic
>> shutdown without mentioning the word "finalize", but that ship has
>> sailed. We did specify it and people can expect it to be usable and
>> they can expect it to work. While encouraging people to move away
>> from finalization is a good thing you have to give them a workable
>> replacement.
>>
>>> The initiative is to identify and remediate existing uses of
>>> finalization in the JDK.
>>> The primary concern is about subclasses that reply on the current spec.
>>> If I'm using grepcode correctly[2], it does not show any subclasses
>>> of ThreadPoolExecutor that
>>> override finalize; so it may be a non-issue.
>>
>> Pardon my skepticism if I don't consider "grepcode" as a reasonable
>> indicator of whether there are subclasses of TPE that override
>> finalize. Only a small fraction of source code is in the open.
>>
>> And I have to agree with Martin that the current documentation for
>> finalize() in TPE is somewhat inappropriate. If you deprecate
>> something you're supposed to point the user to the preferred way of
>> doing something other than the deprecated method - but there is none.
>> finalize() in TPE should not have been deprecated until we provided
>> that alternative.
>>
>> I think the way forward here would be to:
>>
>> 1. Change TPE.finalize() to final to force any subclasses to migrate
>> to the new mechanism going forward.
>>
>> 2. Implement the new mechanism - presumably Cleaner - and document
>> how to achieve the effect of "override to specialize then call super"
>> (presumably by overriding shutdown() instead).
>>
>> then in a few releases we remove TPE.finalize() (at the same time as
>> we remove it from Object).
>>
>> Cheers,
>> David
>>
>>> Regards, Roger
>>>
>>> [1] https://bugs.openjdk.java.net/browse/JDK-8165641
>>>
>>> [2]
>>> http://grepcode.com/search/usages?type=method&id=repository.grepcode.com%24java%24root@jdk%24openjdk@8u40-b25@java%24util%24concurrent@ThreadPoolExecutor@finalize%28%29&k=d
>>>
>
More information about the core-libs-dev
mailing list