I think the dmb instructions are correct even for the non-coherent case

In D60099#1379536, @andrew wrote:

I think the dmb instructions are correct even for the non-coherent case

Trying to determine what you mean here.. Are you saying do the dmb for everyone, return on a NULL map, and continue to do the dsb() for non-coherent / bouncy devices?

You can replace the existing dsb calls in with the appropriate dmb from the map == NULL block.

In D60099#1380228, @andrew wrote:

You can replace the existing dsb calls in with the appropriate dmb from the map == NULL block.

From what I've read, I'm not sure that dmb is enough for non-coherent devices. At least it seems risky. I'd strongly prefer not to do that, as I don't have a non-coherent device that I can test with.
Would you be OK with the patch as-is? That's the limit of what I can test and what I feel comfortable with

I talked with the Arm Linux IOMMU/DMA maintainer about this & we concluded replacing the dsb's with dmb's are correct. I have HW & a local patch that I can test next week.

I've created D60407 that should handle both coherent & non-coherent cases