Hi, Christoph,
It would be good to revisit this to see if we could set a direction for supporting user-space mapping of dma-coherent memory with TTM.
I think in the end, what TTM needs is a DMA interface similar to the one outlined here:
https://lore.kernel.org/lkml/b811f66d-2353-23c6-c9fa-e279cdb0f832@shipmail.o...
That could without too much effort be implemented for all architectures TTM supports and also for those weird architectures you've described that need special treatment, it would probably suffice with a more elaborate definition of the dma_pfn_t together with the dma coherent fault work you were suggesting earlier.
Regardless, once this patchset is in,
https://lore.kernel.org/lkml/20200304114527.3636-1-thomas_os@shipmail.org/
a short term solution would be to export the dma_pgprot() function and have TTM use it to set the pgprot for dma-coherent memory.
What do you think about this?
Thanks,
Thomas