On Friday 08 July 2011 09:36:19 Jonas Bonn wrote: > I put a bit more thought into what we actually need for DMA and came up > with the implementation in this patch. This is a lot simpler than what > we had before, but it seems to fit nicely with the model that we have. > > The jist of it is that allocation of coherent memory just needs to > return pages with the cache-inhibit bit set so we now: > > i) call __get_free_pages (via alloc_pages_exact) > ii) set the CI bit on those pages > iii) flush the pages from cache > > This should mean that we don't have any conflicting page table entries > for these pages anymore. > > I'd appreciate if you could look over this implementation to make sure > I haven't missed anything fundamental. Looks ok to me. A few more points for future optimization though: * If your CPU can do different page sizes, you should make the kernel mapping use the largest possible page size. This will however make the dma mapping part much harder. * If you ever want to map a frame buffer with this, you should add a write-combining mapping type, uncached accesses are really slow. * Cache coherency is good. If you can make the hardware do cache-coherent DMA, then just use cacheable pages for alloc_coherent. Arnd -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html