Re: Confused by Lua garbage collection tear down order when there are references between objects

[Eric Wing <ewmailing@...>]

On 5/22/14, Coda Highland <chighland@gmail.com> wrote:
> On Thu, May 22, 2014 at 12:36 PM, Eric Wing <ewmailing@gmail.com> wrote:
>> In this example:
>> - I create objectZ, objectA, and objectB (in that order).
>> - objectA keeps a strong reference to objectB and objectZ.
>> - I remove references to objectB and objectZ, proving objectA is
>> keeping them alive.
>> - I remove my reference to objectA and watch what garbage collection
>> does.
>>
>> I expect that objectA is "marked for finalization" first because it is
>> keeping objectB and objectZ alive.
>> Then objectB and objectZ may be "marked for finalization" (no order on
>> these is expected).
>> So "in the reverse order they were marked for finalization", I expect
>> the finalization order to be:
>> 1) objectZ or objectB
>> 2) objectB or objectZ
>> 3) objectA
>
> Your behavior is derived from expectations of a reference-counted
> garbage collector. Lua doesn't refcount. It's mark-and-sweep -- as a
> high-level overview, it maintains a list of all objects, and a list of
> root objects (mostly the top-level _ENV and the registry). When it
> does a sweep, it starts from the root objects and marks every object
> that it can reach as accessible. Then it sweeps over the list of all
> objects and flags every object that isn't marked as accessible for
> finalization. Then it runs finalizers on the objects thus marked, in
> reverse of the order that they were marked. On the next sweep, if the
> object is marked as both "finalized" and "inaccessible" it's deleted.
> (If it's marked as "finalized" but NOT marked as "inaccessible" it's
> been resurrected. It'll never have the finalizer run again.)
>
> /s/ Adam
>

I'll cede that my brain is probably poisoned by reference counting.
However, if I cede that point, from your description of the process,
it is still not clear to me "marked for finalization" order is
anything but implementation dependent and is something you can write a
spec to.

If you are walking/sweeping for unreachable objects (and without
regard to the strong references in my original example), then at best,
we seem back to "creation order", or worse "implementation order". I'm
not seeing how the "marked for finalization order" is determined,
standardizable, and distinct from "creation order".


Thanks,
Eric
-- 
Beginning iPhone Games Development
http://playcontrol.net/iphonegamebook/