Re: [Boost-users] [ptr_container]delete_clone and shared libraries

Subject: Re: [Boost-users] [ptr_container]delete_clone and shared libraries
From: Thorsten Ottosen (thorsten.ottosen_at_[hidden])
Date: 2009-06-25 10:30:59

Hi Robert,

Robert Dailey skrev:
> On Windows, I plan to create a ptr_vector in a separate DLL and 
> propagate it to the executable for ownership (Factory-style). The 
> executable will be responsible for the lifetime of this ptr_vector, and 
> will eventually result in "delete" being called on each element in the 
> ptr_vector. However, this 'delete' will be from the executable's memory 
> manager, when the memory was originally allocated in the shared 
> library's memory manager, causing serious issues.
> 
> It would be ideal to specify a delete_clone class that ptr_vector would 
> use (via ADL) and would force those objects in the ptr_vector to use the 
> 'delete' call from the DLL's memory manager, and not the executable's.
> 
> I don't see that this is going to be possible, or at least able to be 
> done easily. Since the delete_clone's operator will have to be 
> implemented in a CPP file, so it will be included into the shared 
> library. This also implies that I'll have to import the delete_clone's 
> interface, since this DLL will be loaded dynamically with LoadLibrary() 
> in Windows. I just don't see this happening.
> 
> Now, if ptr_vector had a function like so:
> 
> ptr_vector::set_deleter()
> 
> I could give it a pointer to a global function in my DLL, and the 
> deleter becomes fully transparent to my executable. It just has to keep 
> the ptr_vector around and let it fall out of scope and the deleter from 
> the DLL will be used to delete each object in the ptr_vector. However, 
> it isn't this simple and I need advice on how I can make this work.

Well, we need to make the clone_allocator stateful, which is currently 
not supported. I always was lacking a good argument, so let's work on 
the design of this on the dev list.

-Thorsten