[openib-general] NFS/RDMA client release for Linux 2.6.15

[Tom Tucker]

Are we all talking about the same thing here? I think Christoph is just
asking that the code use dma_map_[single|page|sg] instead of using
page_to_phys or virt_to_phys.

The rest of this talk is about memory registration strategies which is a
different issue than having the code assume that bus address == phys
addresses.

> On Mon, 2006-02-20 at 10:12 -0800, Caitlin Bestler wrote:
> 
> 
> On 2/20/06, Roland Dreier <rdreier at cisco.com> wrote:
>             Thomas> And, this is only one of many memory registration
>             Thomas> modes. We would use memory windows, if only OpenIB
>             Thomas> provided them (yes I know the hardware currently
>         sucks for
>             Thomas> them). We will add FMR support shortly. In both
>         these 
>             Thomas> modes we perform all addressing by the book via
>         1-1 OpenIB
>             Thomas> registration.
>         
>         Memory windows don't solve anything here, do they?  You still
>         have to
>         register the full region using bus addresses somehow. 
>         
>         - R.
>         _______________________________________________


> Keep in mind that there are two problems: registering memory and
> exposing memory.
> Windows solves the latter problem. FMR work requests solves both.
> 
> A kernel based storage related client will frequently want to form
> logical buffers
> from scattered physical pages. The pages selected are not necessarily
> part
> of an existing virtual memory map, and especially not a registered
> one.
> 
> You can solve the memory registration problem once by creating an
> exportable memory region that covers all of physical memory. The
> problem is that you do not want to advertise that regions RKey/STag.
> 
> Memory Windows solve that problem, by allowing you to bind windows
> within the memory region. The problem is that if the buffer is not
> physically
> continquous then you still have to export a multi-element list in
> order to
> have the peer to read from a discontiguous target.
> 
> A Fast Memory Region allows arbitrary sets of pages to form a single
> logical window for the purposes of peer-to-peer interaction, and has a
> life cycle that more naturally maps the duration when the pages have
> to be iomapped.
> 
> 
> 
> 
> 
> _______________________________________________
> 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