On Fri, May 31, 2019 at 06:45:00PM +0100, Robin Murphy wrote:

Bleh, I'm certainly not keen on formalising any kind of
dma_to_phys()/dma_to_virt() interface for this. Or are you just proposing
something like dma_unmap_sorry_sir_the_dog_ate_my_homework() for drivers
which have 'lost' the original VA they mapped?

Yes, I guess we need that in some form.  I've heard a report the IBM
emca ethernet driver has the same issue, and any SOC with it this
totally blows up dma-debug as they just never properly unmap.