Well yes - when you OBJ_RELEASE(buf), it calls free on the data blob. All you
have to do is set buf->base_ptr = NULL prior to releasing.
You could also save a lot of malloc wastage by just using
OBJ_CONSTRUCT/DESTRUCT for buf, though you still have to NULL the base_ptr
before calling DESTRUCT
> On Feb 11, 2015, at 2:57 PM, Howard Pritchard <hpprit...@gmail.com> wrote:
>
> Hello Folks,
>
> I've got a question about the opal_dss.load functionality.
>
> If I have a construct like
>
>
> buf = OBJ_NEW(opal_buffer_t);
> opal_dss.load(buffer, input_ptr, num_bytes_to_load);
>
> OBJ_RELEASE(buf);
>
> I'm finding that input_ptr has to have been something
> returned by malloc and friends or else I get heap corruption.
>
> I can't do something like
>
> oob_allgatherv(input_ptr, bytes_from_each_rank, output_ptr);
>
> for (cptr = output_ptr, i = 0 ; i < pmix_size ;i ++) {
> opal_dss.load(buf, cptr, bytes_from_each_rank[i]);
> unpack message into dstore
> cptr += bytes_from_each_rank[i];
> }
>
> I have to create a separate malloc'd buffer each time through the
> loop and do a memcpy into the tmp buffer, then call opal_dss.load
> for that particular tmp buffer.
>
> Is this expected behavior of opal_dss.load? Is there a way to
> avoid these extra mallocs/memcopy's?
>
> If I don't do the above technique, the heap allocator blows up
> in OBJ_RELEASE of buffer.
>
> Thanks,
>
> Howard
>
>
>
> _______________________________________________
> devel mailing list
> de...@open-mpi.org
> Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel
> Link to this post:
> http://www.open-mpi.org/community/lists/devel/2015/02/16968.php
