On Wed, 2014-10-01 at 17:41 +1000, Benjamin Herrenschmidt wrote: > I don't like that approach because technically the chip doesn't do > 64-bit DMA ... it does something like 40 or 48 (might actually depend on > the chip version) and for all I know it will break on future x86 with > more memory or other platforms with similar address encodings as > powerpc... > > The right thing might be to get the exact number of bits and do the > appropriate dma_set_mask() like the graphics driver does, but that's a > bit tricky unless we add a DMA mask field in that big array of chips in > there... Note that I made the assumption that in the grand scheme of things, HDMI audio isn't a very big throughput device and thus the extra iommu mappings (or swiotlb) shouldn't hurt that much. But I may be wrong... So by all means, let me know what you want done here, but I don't think CONFIG_PPC is the right thing to do here, especially since this is not even something that affects all PPCs, but specifically the POWER server PCI bridges for which we configure a 64-bit window high up in the address space. FYI: Basically we setup PCI DMA addresses 0...2G as a 32-bit IOMMU remapped window and we also configure a "direct map" window higher up that bypasses the IOMMU and provides a direct linear mapping of RAM for 64-bit capable devices. The problem is that the HW on those server chips requires that second window to have address bit 59 set. Then based on the DMA mask, the arch code provide the device with DMA ops that will either go through the iommu remap window or through the bypass window based on the DMA mask set by the driver. Cheers, Ben. -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html