[boost] Re: Weak ref. via atomicity (RE: Smart pointers:...)

From: Alexander Terekhov (terekhov_at_[hidden])
Date: 2003-02-13 12:03:09

Pavel Vasiliev wrote:
[...]
> > pthread_refcount_decrement()         // with msync for proper mut.obj-cleanup

Basically [in terms of proposed/revised Java MM], it's VOLATILE-RELEASE 
when count > 1 'prior to decrement' and VOLATILE-ACQUIRE when the count
drops to 0. I think that one might even try to optimize-away VOLATILE-
RELEASE msync [and interlocked instruction(s)] for the'count == 1'-case; 
on some platforms, perhaps.

> > pthread_refcount_decrement_nomsync() // without any msync whatsoever
> 
> "nomsync" here stands for "no memory view update even when counter drops
> to 0"? 

Yes, it's a variant without [VOLATILE-RELEASE/VOLATILE-ACQUIRE] msync. 

> If yes, then you probably should not use it in
> 
> >  void release_weak() {
> >    int status = pthread_refcount_decrement_no_msync(&weak_count);
> >    if (PTHREAD_REFCOUNT_DROPPED_TO_ZERO == status)
> >      destruct_self();
> >  }
> 
> Consider the following:
> > Given: two threads -- A and B,
> >        thread A has "strong" one,
> >        thread B has "weak" one,
> >        strong_count == 1, weak_count == 2.
> 
> Thread A releases the last "strong" reference.
> destruct_object() deletes p_object AND modifies the refs instance
> (e.g. sets p_object to NULL for debug reasons).

Yeah... "cleanup" for intrusively counted objects [with refs embedded 
within the object] aside for a moment...

> 
> strong_count == 0, weak_count == 1
> 
> Thread B releases the last "weak" reference.
> release_weak() calls destruct_self() on expired memory view.
> 
> "Or am I just missing and/or misunderstanding something?" :-)

No, you're right. Well, let me put it this way: < ;-) >

  Surely you have missed something! See the "// with msync for..." 
  comment above. It says "for proper mut.obj-cleanup". "mut." stands 
  for mutable. Obviously, the fact that _no_msync() variant is used
  here indicates that destruct_object() [and the rest too] is meant 
  to be >>'const'<< qualified method[s] that shall not modify any 
  >>'mutable'<< members (things like "pthread_refcount_t" counts, 
  mutexes, if any). And all this was crystal clear, I mean.

Good catch. Thanks.
     ^^^^^(*) Apropos...

regards,
alexander.

(*) http://groups.google.com/groups?selm=3E4B57AB.8BBF5A32%40web.de
    (Subject: Re: exceptions)