12:30 <cyndis> tagr: How was the limitation of IOVAs to 39 bits supposed to work on Tegra194? I'm only seeing your device tree patch to add the memory-controller node with dma-ranges

12:30 <cyndis> is there perhaps some magic in the interconnects?

12:31 <cyndis> the problem I'm seeing is that with context devices, I'm getting full 40 bit IOVAs currently

12:31 <cyndis> they are not platform devices so if the restriction is coming though device tree, that is not surprising

12:32 <tagr> yeah, there's some magic in the core OF code that sets the bus' DMA limit using dma-ranges

12:33 <cyndis> MC is not a parent of all of our devices though so I wasn't expecting that to apply

12:34 <tagr> cyndis: of_dma_configure_id() is where this happens

12:34 <tagr> it's a DMA parent of all the device, though

12:35 <tagr> that relationship is created via the interconnects property, and that's also why the naming is a bit inconsistent in some places, because in order for this to work, there has to be one interconnect path named "dma-mem"

12:35 <cyndis> I see, so it's interconnects magic

12:37 <cyndis> I do use of_dma_configure_id for the context devices so I guess if I add an interconnects property it probably will work

12:37 <cyndis> though we actually need to limit it to 38 bits since NVENC has a bug where it cannot use the last bit..

12:38 <cyndis> also the context devices are not real devices so they are not really connected to the memory-controller

12:38 <cyndis> maybe I'll just set the DMA mask directly

17:01 <tagr> they are connected via their parent, right, so perhaps just inherit the DMA mask from the parent, which I guess is the host1x device

17:03 <tagr> but yeah, if we need to special-case for NVENC anyway, might as well set it explicitly (or you could perhaps do it conditionally for only the NVENC context device)

17:40 <cyndis> the issue is that there is no "NVENC context device", since each process gets a context and we don't know if they're going to use nvenc or not

17:41 <cyndis> but, this seems to work, so I think it's good