On Tue, Dec 18, 2018 at 07:50:12PM +0100, Christoph Hellwig wrote: > On Tue, Dec 18, 2018 at 11:43:08AM -0700, Jason Gunthorpe wrote: > > So the problem here is that on some arches > > > > sg_set_buf(sg, dma_coherent_buf, size) > > p = sg_virt(sg); > > assert(p == dma_coherent_buf); > > dma allocations purely return a virtual address, you must never > call virt_to_page or virt_to_phys on them, which sg_set_buf > will do. On many architectures this will give your the wrong > result as the coherent DMA address is a vmap or ioremap address. Yes, that is what I gathered - if that is the design then I'd say that drivers shouldn't be stuffing these DMA coherent virtual addresses into a sg at all. Jason
