[openib-general] Re: [PATCH][RFC][0/4] InfiniBand userspace verbs implementation

[Caitlin Bestler]

That leaves a problem when the same memory region is
registered with different vendors. Verbs A marks the area,
Verbs B sees that is already marked, Verbs A unmarks
the area when it is done not knowing that B is relying
on the memory staying pinned.

I do not believe there is a solution to this problem when
working at arms length from Linux other than documenting
the problem and informing applications of workarounds
required when using multiple vendors concurrently with
the same memory (i.e, destroy the most recently
created memory region first, or pin the memory
yourself before creating the first memory region).

The only other alternative is to make the pinning
some sort of shared service that would apply across
multiple vendors. That is doable, but might not be
worthwhile given that a single process using multiple
vendor devices concurrently is decidely the exception.
But those users deserve at least a warning.


On 4/25/05, Roland Dreier <roland at topspin.com> wrote:
>     Caitlin> Who is responsible for counting within a process, and
>     Caitlin> then between processes (in case shared memory is being
>     Caitlin> registered)? The application? Middleware? Driver?
> 
> The verbs code doing the registration should do it as part of the
> registration.  Shared memory does not cause any additional issues
> because it is mapped into the virtual memory map of each process and
> must be marked VM_DONTCOPY in each process separately.
> 
>  - R.
> 
> _______________________________________________
> openib-general mailing list
> openib-general at openib.org
> http://openib.org/mailman/listinfo/openib-general
> 
> To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
>