Hi all, Thanks for your comments. On 2016-03-11 03:15:22 -0800, Christoph Hellwig wrote: > On Thu, Mar 10, 2016 at 10:47:10PM -0800, Dan Williams wrote: > > I think it is confusing to use the dma_ prefix for this peer-to-peer > > mmio functionality. dma_addr_t is a device's view of host memory. > > Something like bus_addr_t bus_map_resource(). Doesn't this routine > > also need the source device in addition to the target device? The > > resource address is from the perspective of the host cpu, it may be a > > different address space in the view of two devices relative to each > > other. > > Is it supposed to be per-mmio? It's in dma-mapping ops, and has dma > in the name, so I suspected it's for some form of peer dma. But given > that our dma APIs reuqire a struct page backing I have no idea how this > even supposed to work, and this little documentation blurb still doesn't > clear that up. > > So for now I'd like to NAK this patch until the use case can be > explained clearly, and actually works. I can explain the use case and maybe we can figure out if this approach is the correct one to solve it. The problem is that I have devices behind an IOMMU which I would like to use with DMA. Vinod recently moved forward with his and Linus Walleij patch '[PATCH] dmaengine: use phys_addr_t for slave configuration' which clarifies that the DMA slave address provided by a client is the physical address. This puts the task of mapping the DMA slave address from a phys_addr_t to a dma_addr_t on the DMA engine. Without an IOMMU this is easy since the phys_addr_t and dma_addr_t are the same and no special care is needed. However if you have a IOMMU you need to map the DMA slave phys_addr_t to a dma_addr_t using something like this. Is it not very similar to dma_map_single() where one maps processor virtual memory (instead if MMIO) so that it can be used with DMA slaves? -- Regards, Niklas Söderlund
