Hi, Arnd Bergmann writes: > On Thursday, September 1, 2016 5:14:28 PM CEST Leo Li wrote: >> >> Hi Felipe and Arnd, >> >> It has been a while since the last response to this discussion, but we >> haven't reached an agreement yet! Can we get to a conclusion on if it >> is valid to create child platform device for abstraction purpose? If >> yes, can this child device do DMA by itself? > > I'd say it's no problem for a driver to create child devices in order > to represent different aspects of a device, but you should not rely on > those devices working when used with the dma-mapping interfaces. heh, that looks like an excuse to me :-) This will always be a problem for e.g. MFD, for example. Are you saying MFD child-devices shouldn't be allowed to do DMA? It becomes silly when you read it that way, right? > This used to be simpler back when we could configure the kernel for > only one SoC platform at a time, and the platforms could provide their > own overrides for the dma-mapping interfaces. These days, we rely on right, so we have a very old regression that just took a complex driver such as dwc3 to trigger ;-) > firmware or bootloader to describe various aspects of how DMA is done, there's no DMA description in DT. Every OF device gets the same 32-bit DMA mask and that is, itself, wrong for several devices. > so you can't assume that passing a device without an of_node pointer > or ACPI data into those functions will do the right thing. That's not the problem, however. We can very easily pass along ACPI_COMPANION() to any platform_device we want, but that's not enough because DMA-related bits are passed along with archdata; but archdata isn't generic in any way. Some arches (like x86) _do_ use it for DMA, but some don't. -- balbi