Hi Thierry, On Tue, Jan 30, 2018 at 11:01:50AM +0100, Thierry Reding wrote: > On Tue, Jan 30, 2018 at 10:59:16AM +0100, Thierry Reding wrote: > > On Tue, Jan 30, 2018 at 10:24:48AM +0100, Arnd Bergmann wrote: > > > On Tue, Jan 30, 2018 at 8:54 AM, Maxime Ripard > > > <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote: > > > > On Mon, Jan 29, 2018 at 03:34:02PM +0100, Arnd Bergmann wrote: > > > >> On Mon, Jan 29, 2018 at 10:25 AM, Linus Walleij > > > >> <linus.walleij@xxxxxxxxxx> wrote: > > > >> > On Mon, Jan 29, 2018 at 9:25 AM, Maxime Ripard > > > >> > <maxime.ripard@xxxxxxxxxxxxxxxxxx> wrote: > > > >> >> On Sat, Jan 27, 2018 at 05:14:26PM +0100, Linus Walleij wrote: > > > > > > >> > > > >> At one point we had discussed adding a 'dma-masters' property that > > > >> lists all the buses on which a device can be a dma master, and > > > >> the respective properties of those masters (iommu, coherency, > > > >> offset, ...). > > > >> > > > >> IIRC at the time we decided that we could live without that complexity, > > > >> but perhaps we cannot. > > > > > > > > Are you talking about this ? > > > > https://elixir.free-electrons.com/linux/latest/source/Documentation/devicetree/bindings/dma/dma.txt#L41 > > > > > > > > It doesn't seem to be related to that issue to me. And in our > > > > particular cases, all the devices are DMA masters, the RAM is just > > > > mapped to another address. > > > > > > No, that's not the one I was thinking of. The idea at the time was much > > > more generic, and not limited to dma engines. I don't recall the details, > > > but I think that Thierry was either involved or made the proposal at the > > > time. > > > > Yeah, I vaguely remember discussing something like this before. A quick > > search through my inbox yielded these two threads, mostly related to > > IOMMU but I think there were some mentions about dma-ranges and so on as > > well. I'll have to dig deeper into those threads to refresh my memories, > > but I won't get around to it until later today. > > > > If someone wants to read up on this in the meantime, here are the links: > > > > https://lkml.org/lkml/2014/4/27/346 > > http://lists.infradead.org/pipermail/linux-arm-kernel/2014-May/257200.html > > > > From a quick glance the issue of dma-ranges was something that we hand- > > waved at the time. > > Also found this, which seems to be relevant as well: > > http://lists.infradead.org/pipermail/linux-arm-kernel/2014-May/252715.html > > Adding Dave. Thanks for the pointers, I started to read through it. I guess we have to come up with two solutions here: a short term one to address the users we already have in the kernel properly, and a long term one where we could use that discussion as a starting point. For the short term one, could we just set the device dma_pfn_offset to PHYS_OFFSET at probe time, and use our dma_addr_t directly later on, or would this also cause some issues? For the long term plan, from what I read from the discussion, it's mostly centered around IOMMU indeed, and we don't have that. What we would actually need is to break the assumption that the DMA "parent" bus is the DT node's parent bus. And I guess this could be done quite easily by adding a dma-parent property with a phandle to the bus controller, that would have a dma-ranges property of its own with the proper mapping described there. It should be simple enough to support, but is there anything that could prevent something like that to work properly (such as preventing further IOMMU-related developments that were described in those mail threads). Maxime -- Maxime Ripard, Free Electrons Embedded Linux and Kernel engineering http://free-electrons.com
Attachment:
signature.asc
Description: PGP signature
