[openib-general] [PATCH] [RMPP] receive RMPP support

[Grant Grundler]

On Fri, Mar 18, 2005 at 08:15:04PM -0800, Roland Dreier wrote:
> it's somewhat risky to use dma_map_single() on fields in the middle of
> a structure, because you don't know that the field starts and ends on
> a cacheline boundary.  In this case you're pretty safe because you're
> doing DMA_TO_DEVICE, but if you use the same type of code with
> DMA_FROM_DEVICE on a non-cache-coherent arch (e.g. PowerPC 4xx) then
> you can get into trouble.  See <http://lwn.net/Articles/2265/> for a
> very nice writeup of the problem.

Indeed. I never saw even that though I review DMA_mapping.txt
and wrote the DMA support for parisc-linux port.
Good catch and have to wonder if PCI pool is really the right
choice (or not) for inbound DMA....need some way of enforcing
that in order to catch it before it's a problem.

thanks,
grant