[squeak-dev] Re: Multiple finalizers per single object

Levente Uzonyi leves at elte.hu
Fri Sep 24 13:00:04 UTC 2010


On Fri, 24 Sep 2010, Igor Stasenko wrote:

> 2010/9/24 Levente Uzonyi <leves at elte.hu>:
>> On Fri, 24 Sep 2010, Igor Stasenko wrote:
>>
>>> 2010/9/24 Levente Uzonyi <leves at elte.hu>:
>>>>
>>>> On Fri, 24 Sep 2010, Igor Stasenko wrote:
>>>>
>>>>> On 24 September 2010 01:40, Andreas Raab <andreas.raab at gmx.de> wrote:
>>>>>>
>>>>>> On 9/23/2010 3:27 PM, Igor Stasenko wrote:
>>>>>>>
>>>>>>> On 23 September 2010 23:58, Andreas Raab<andreas.raab at gmx.de>  wrote:
>>>>>>>>
>>>>>>>> Isn't that a bit of a made-up problem? I've never seen (nor heard of)
>>>>>>>> anyone
>>>>>>>> even trying to attach more than one finalizer to an object.
>>>>>>>> Disallowing
>>>>>>>> it
>>>>>>>> could be done, but unless there's a good reason for it I'm not in
>>>>>>>> favor
>>>>>>>> of
>>>>>>>> deciding what is good policy and what isn't without a practical use
>>>>>>>> case
>>>>>>>> at
>>>>>>>> hand. If you have a use-case I'd like to hear more about it, but if
>>>>>>>> not,
>>>>>>>> then *shrug*, who cares :-)
>>>>>>>>
>>>>>>>
>>>>>>> The practical use, that one may mistakenly add two same finalizers on
>>>>>>> a same object,
>>>>>>> like closing a file handle, or free external memory.
>>>>>>> And it is potentially dangerous, because OS could reuse the same
>>>>>>> memory location or file handle,
>>>>>>> once you closed it, so, by closing it twice you can get a serious
>>>>>>> trouble finding what's going on.
>>>>>>>
>>>>>>> My point, that if one ever needs a composite finalization action, a
>>>>>>> WeakRegistry is wrong place for that.
>>>>>>> WeakRegistry should prevent registering more than a signle finalizer
>>>>>>> per unique object to prevent
>>>>>>> potential problems which could arise from such composition(s).
>>>>>>
>>>>>> So you're saying one *could* do that and *if* one did that it *might*
>>>>>> lead
>>>>>> to a *potential* issue? Sounds like solving a non-problem to me :-)
>>>>>>
>>>>>
>>>>> Yes i consider it as a potential security hole. It simply feels wrong to
>>>>> me.
>>>>> It is good that nobody using multiple finalizers.
>>>>
>>>> How do you know that for sure?
>>>> Currently our announcement framework uses this feature to remove all weak
>>>> subscriptions for an object when it's garbage collected. This avoids
>>>> iterating through all the subscriptions looking for those which don't
>>>> have a
>>>> subscriber anymore. I'm sure it could be implemented another way, but
>>>> it's
>>>> really simple this way.
>>>>
>>> This is userful.
>>> Though, iterating through all subscriptions, this is what you can
>>> avoid using new finalization scheme.
>>>
>>> When weak field of subscription becomes nil, a subscription will be
>>> automatically added to the list of expired subscriptions.
>>> Then, after GC, you simply go through that list and remove them from
>>> subscriptions container.
>>> Note that such implementation doesn't involving WeakRegistry at all,
>>> and therefore you won't be needing
>>> multiple finalizers per single object.  :)
>>
>> I think we can live without multiple finalizers per object.
>> Maybe the best solution is to make the behavior changeable when an object
>> already has a finalizer in a WeakRegistry. With the current WeakRegistry
>> this could be something like:
>>
>> WeakRegistry >> add: anObject executor: anExecutor
>>        "Add anObject to the receiver. Store the object as well as the
>> associated executor."
>>
>>        self protected: [
>>                (valueDictionary associationAt: anObject ifAbsent: nil)
>>                        ifNil: [ valueDictionary at: anObject put: anExecutor
>> ]
>>                        ifNotNil: [ :association |
>>                                multipleFinalizerPerObjectAction
>>                                        cull: association
>>                                        cull: anExecutor ] ]
>>
>> WeakRegistry >> initialize: n
>>
>>        valueDictionary := WeakIdentityKeyDictionary new: n.
>>        accessLock := Semaphore forMutualExclusion.
>>        multipleFinalizerPerObjectAction := [ :association |
>>                self error: association key printString, ' already has an
>> executor!' ].
>>        self installFinalizer.
>>
>> WeakRegistry >> multipleFinalizerPerObjectAction: aBlock
>>
>>        multipleFinalizerPerObjectAction := aBlock
>>
>>
>> This way one could
>> - have multiple finalizers per object as it is in Squeak 4.1
>> - raise an error as you suggested
>> - replace the current executor with the new one (as it was before Squeak
>> 4.1).
>>
>> I don't know if it fits your new WeakRegistry implementation. I'm also not
>> sure if this isn't a bit over-engineered.
>>
>
> I had to resemble same functionality, as in squeak 4.1 ,  so new
> registry still supports
> multiple finalizations per object.
> Please, try to file-in the code in http://bugs.squeak.org/view.php?id=7473
> and see if you have any problems with it.

I did that this morning. Here's what I found:

These can be loaded in one shot (the order doesn't matter):
phase-0.cs
phase-1-weakfinalizationlist.st
phase-1-weakfinalizationregistry.st
phase-1-weakfinalizeritem.st
phase-4-weakfinalizerstest.st

This can be the second step, it can be merged with the next one:
phase-2-finalizationprocess.1.cs

This can be the thrid step, it could be a MC postscript:
phase-3-migrating.1.cs

I think the migration code should be removed when the integration is done.

After installing the code I tried to explore [WeakRegistry allInstances], 
but got a debugger, because some WeakRegistry instances didn't have their 
sema variable initialized. After a GC this issue was solved.

When I was looking at the changes compared to the current code I found 
that WeakRegistry >> #keys is not protected in the new version, which may 
not return all the keys in a rare case.

With the current migration technique we lose the history of WeakRegistry. 
I think it's possible to do the migration without losing the history by 
simply extending the current class with the new methods (using different 
names if they already exist), adding the list variable to the current 
class, then migrating the instances, renaming the clashing methods, 
finally removing the obsolete stuff.

Also a few comments are lost during the migration, those should also 
restored IMHO.

Today or tomorrow I'll check the code more in-depth and prepare mczs from 
your changesets (if noone else does it before me).


Levente

>
>>
>> Levente
>>
>>>
>>>>
>>>> Levente
>>>>
>>>>> Okay, then why they are here? Code should have a practical use,
>>>>> otherwise it is just a garbage, no?
>>>>> Or is it an invitation to start using them?
>>>>> I'm not envy to everyone who will start doing it. :)
>>>>>
>>>>> I'd really like to see a good practical example where such thing could
>>>>> be userful,
>>>>> which outweights a potential problems i described before.
>>>>>
>>>>>> (but go ahead and fix it if you feel that it needs fixing)
>>>>>>
>>>>>> Cheers,
>>>>>>  - Andreas
>>>>>>
>>>>>>>
>>>>>>>> Cheers,
>>>>>>>>  - Andreas
>>>>>>>>
>>>>>>>> On 9/23/2010 1:23 PM, Igor Stasenko wrote:
>>>>>>>>>
>>>>>>>>> Hello,
>>>>>>>>>
>>>>>>>>> i'd like to raise this subject once more, because i don't like it
>>>>>>>>> (or
>>>>>>>>> don't understand?).
>>>>>>>>>
>>>>>>>>> In all scenarios, where i met the need to use finalization, a single
>>>>>>>>> finalizer is sufficient.
>>>>>>>>> Moreover, there is always a single object who controls a weak
>>>>>>>>> reference, and it never leaks it out, to prevent
>>>>>>>>> the case, when some other object may obtain a strong reference on
>>>>>>>>> it,
>>>>>>>>> making it permanently held in object memory.
>>>>>>>>>
>>>>>>>>> Multiple different finalizers for single object, from design point
>>>>>>>>> of
>>>>>>>>> view, means that you having two different, not related frameworks,
>>>>>>>>> which using same object, and want to do something when it dies.
>>>>>>>>> A scenario, where its possible and userful, still don't comes into
>>>>>>>>> my
>>>>>>>>> mind.
>>>>>>>>> In contrary, whenever i see a use of finalizers, its in most cases
>>>>>>>>> about graceful control over external resource, such as:
>>>>>>>>> - file
>>>>>>>>> - socket
>>>>>>>>> - external memory
>>>>>>>>>
>>>>>>>>> and i really don't see how multiple finalizers per single resource
>>>>>>>>> could do any good.
>>>>>>>>>
>>>>>>>>> Suppose one finalizer closing a file handle, while another one
>>>>>>>>> flushing it buffer cache.
>>>>>>>>> Now, how you going to ensure, that one finalizer will execute first,
>>>>>>>>> before another one?
>>>>>>>>> And what if third framework comes into play and wants to add another
>>>>>>>>> finalizer on top of that, which should do something in the middle
>>>>>>>>> between flushing a cache and closing file handle?
>>>>>>>>>
>>>>>>>>>>  From the above, the only conclusion can be made: use a single
>>>>>>>>>
>>>>>>>>> finalizer, and put all logic&    operation ordering into it.
>>>>>>>>> And also, prevent leakage of object pointer (such as file handle)
>>>>>>>>> outside of your model, otherwise it may cause harm.
>>>>>>>>>
>>>>>>>>> That's why i think a current WeakRegistry model provoking bad design
>>>>>>>>> practices.
>>>>>>>>> I think a better behavior would be to raise an error, if something
>>>>>>>>> wants to register finalizer twice for a single object.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Best regards,
>>>>> Igor Stasenko AKA sig.
>>>>>
>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>>
>>> --
>>> Best regards,
>>> Igor Stasenko AKA sig.
>>>
>>
>>
>>
>>
>
>
>
> -- 
> Best regards,
> Igor Stasenko AKA sig.
>
>


More information about the Squeak-dev mailing list