Preparation for unpinned DMA-buf handling

[Daniel Vetter]

On Tue, May 14, 2019 at 04:05:13PM +0200, Christian König wrote:
> Hi Daniel,
> 
> since we are not moving forward with this I've separated the change out of the larger patchset.
> 
> Any objections to this now? It's basically a 1:1 move of the functionality from DRM to DMA-buf.

For my big picture understanding, and because I spotted in the latest rocm
release that apparently the fancy interconnect is now support. I looked
around a bit in upstream and basic xgmi support seems to be there already,
but I haven't seen anything that looks like xgmi buffer sharing is there
already? Missing something, or am I looking at the wrong thing (xgmi seems
to be the engineering name for that fancy interconnect at least).

For the big plan I've seen the dynamic dma-buf and p2p stuff, I think it'd
be good if we also make sure xgmi and related things integrate somewhat
neatly. Ofc xgmi requires that we upcast the dma_buf_attachment to
somethign driver specific (I don't think we should make that generic, but
maybe Jerome has different ideas with HMM). Question I have is how much of
the map/unmap/invaliidate and passing sg tables around we can reuse for
this.

If you have patches/pointers would be really intersting to read them a
bit.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch