Re: [boost] Would anybody like a wrapper atomic_shared<T> around shared_ptr<T>?

From: Dean Michael Berris (mikhailberis_at_[hidden])
Date: 2008-07-22 21:24:21

Hi George,

On Wed, Jul 23, 2008 at 6:04 AM, Talbot, George <Gtalbot_at_[hidden]> wrote:
>
> A while back (~2006 I think) I was asking around on this list about
> using boost::shared_ptr<T> to build a composite data structure that can
> be accessed and modified by multiple threads.
>
> Peter Dimov's suggestion at the time was to write a wrapper class for
> boost::shared_ptr<T> that would protect access to the
> boost::shared_ptr<T> with a spinlock.  I've done this, and this works
> pretty good in my application.  I'm interested in contributing this code
> to BOOST, as it's been quite useful.
>
> I have a few questions:
>
>   1)  Is there interest in such a donation.  It's just one header file,
>       and it's wafer-thin...

I am interested, although I do have some reservations with it not
using the locks already available in Boost.Thread.

>   2)  If so, how does one go about making such a donation?

(I think it's a little funny that you use 'donation' instead of
contribution... but nonetheless... ;-) )

>From what I gather, this is something that may have to go through a
mini-review. The veterans can answer this question better though.

>   3)  My implementation below uses a spinlock to protect things and
>       as one might expect, under load my application does spend some
>       time in ::pthread_spin_lock().  Is there a way to rewrite this
>       class (possibly with friendship of boost::shared_ptr<T>) in a
>       lock-free manner?  I've attached the code below...
>

I'm not sure about lock-free approaches, but doesn't shared_ptr<T>
already use atomic operations when built in multi-threaded mode?

Another issue I see is the dependence to ::pthread_spin_lock() -- can
you use a shared recursive_mutex instead?

HTH

-- 
Dean Michael C. Berris
Software Engineer, Friendster, Inc.