On 03/17/2014 06:12 PM, Laurent Pinchart wrote: > Hi Arnd, > > On Friday 14 March 2014 17:57:48 Arnd Bergmann wrote: >> On Friday 14 March 2014, Santosh Shilimkar wrote: >>> I remember NAKing this approach in past and my stand remains same. The >>> cache APIs which you are trying to use here are not suppose to be used >>> outside. >>> >>> I think the right way to fix this is to make use of streaming APIs. >>> If needed, IOMMU can have its own dma_ops for special case handling if >>> any. >>> >>> Russell, Arnd might have more ideas. >> >> I have a bad feeling about using the dma-mapping API within the IOMMU code, >> because that driver is also used to implement the dma-mapping API for >> devices attached to the IOMMU. It's possible that it just works. > > Right, but I hope the memory port used by the IOMMU to fetch page table > entries will not go through itself the same IOMMU :-) > >> Is the IOMMU actually designed to have page tables in noncoherent memory? I >> would have expected that the iopt accesses must all be done on >> dma_alloc_coherent() provided memory to guarantee proper accesses. > > I'm not knowledgeable enough about the OMAP IOMMU, but at least the Renesas > IOMMU doesn't need coherent memory as long as the driver makes sure that > changes to the page tables are made visible to the device. That is the case with OMAP IOMMU as well, we have all along been using non-coherent memory. Switching to coherent memory will probably simplify things in the IOMMU driver (and eliminate this discussion on OMAP :)) regards Suman -- To unsubscribe from this list: send the line "unsubscribe linux-omap" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html