On 11 October 2010 12:10, Stéphane Ducasse <[email protected]> wrote:
> I integrate
>        Issue 3086:     MCMethodDefinition>>shutdown hack to avoid lock down.
>        http://code.google.com/p/pharo/issues/detail?id=3086
>
>
>
> Igor can you have a look and let me know:
>        1- if we should rollback Issue 3086:    MCMethodDefinition>>shutdown 
> hack to avoid lock down.
>        2- if we should integrate
>                Issue 3048:     MC method cache fix
>                http://code.google.com/p/pharo/issues/detail?id=3048
>
> I will integrate
>        http://code.google.com/p/pharo/issues/detail?id=3091
>

Wait , MCMethodDefinition is only a consequence (however, removing it
from weakdependents would be good idea as well ;)).

The real cultprit is HostSystemMenusProxy , which locks the weak
registry during finalization.
Then MCMethodDefinition>>shutdown simply locks at same point (trying
to obtain semaphore lock).
Note, that HostSystemMenusProxy locks finalization process, but not
the process which doing shutdown,
and then during image startup, a new finalization process get created.
That's why removing access to weakdependents in
MCMethodDefinition>>shutdown , kinda solved the issue.
In reality its not.
An old weak registry were sending #finalize outside a critical section:

finalizeValues
        "Some of our elements may have gone away. Look for those and activate
the associated executors."
        | finiObjects |
        finiObjects := nil.
        "First collect the objects."
        self protected:[
                valueDictionary allAssociationsDo:[:assoc|
                        assoc key isNil ifTrue:[
                                finiObjects isNil
                                        ifTrue:[finiObjects := 
OrderedCollection with: assoc value]
                                        ifFalse:[finiObjects add: assoc value]]
                ].
                finiObjects isNil ifFalse:[valueDictionary finalizeValues:
finiObjects asArray].
        ].
        "Then do the finalization"
        finiObjects isNil ifTrue:[^self].
        finiObjects do:[:each| each finalize].


So, there two ways to solve it:
a) perform finalization outside a critical section (see attached)
b) fix the HostSystemMenusProxy to prevent any manupulation with weak
registry during finalization.


I checked Squeak 4.2 image, and i can also say, that any attempt to
manipulate weak registry during finalization will also lock down a
finalization process (see Squeak's WeakRegistry>>finalizeValues &
WeakKeyDictionary>>finalizeValues)

So, if we going to assume, in future that it is error to modify weak
registry during finalization, then we should leave the code in
WeakRegistry as-is, and fix HostSystemMenusProxy instead.
If not, the we should make sure that #finalize should be sent outside
weak registry's critical section.

And the last note: if VM supports new finalization, it will happen
outside a critical section (See the
bottom of WeakRegistry>>finalizeValues). So any manipulations with
registry during finalization can't cause lockdown.


P.S. forwarded to Squeak-dev, to make Squeakers aware of potential
danger , when manipulating weak registry during finalization.

-- 
Best regards,
Igor Stasenko AKA sig.

Attachment: WeakRegistry-finalizeValues.st
Description: Binary data

_______________________________________________
Pharo-project mailing list
[email protected]
http://lists.gforge.inria.fr/cgi-bin/mailman/listinfo/pharo-project

Reply via email to