Ok, I guess this is what Robin was referring to. A midlayer like SCSI really shouldn't directly call dma layer functions without knowing that the underlying bus is DMA capable. I'll see what I can do about it, and in the meantime drop this patch and the companion from the dma-mapping tree.