[Gd-hackers] Suggested patch for Open Dylan's C runtime

[Thomas Christian Chust]

Bruce Hoult wrote:

> On 7/30/07, Hannes Mehnert <hannes at mehnert.org> wrote:
>
>> I just commited that patch to the subversion repository. Thanks a lot.
> 
> This patch appears to put finalized objects onto a queue for later
> running, rather than running them from within the GC. Good!
> 
> BUT .. it does a GC_NEW() from within the finalizer, which is called
> from within the GC.  Therefore the GC is entered recursively.  This is
> very very bad and will bite you one day and you won't know why.

Hello,

I must admit that I don't know the Boehm garbage collector's internals
well enough to decide for sure whether allocation from a finalizer is
safe or not. In case the allocation functions hold some mutex locks this
could indeed break horribly.

But anyway, I think this is about the only way to force the Boehm
garbage collector into an approximately correct behaviour for
finalization queues: If I don't put the object pointer into a queue
allocated with the garbage collector, it gets discarded immediately
after the finalizer has run. Thus I must first make the pointer live
again by putting it into a block of memory allocated with the collector.
Once the queue has been drained, the queue memory itself and the objects
referenced from it will be eligible for collection again, unless the
Dylan finalizers resurrect some objects once more.

The only acceptable alternative I can think of that avoids allocation in
the C-level finalizer would be to preallocate an array list to use as
the finalization queue. But that would pose an artificial size limit on
the finalization queue.

cu,
Thomas