On Fri, Aug 16, 2019 at 02:04:35PM -0700, Rob Clark wrote: > I don't disagree about needing an API to get uncached memory (or > ideally just something outside of the linear map). But I think this > is a separate problem. > > What I was hoping for, for v5.4, is a way to stop abusing dma_map/sync > for cache ops to get rid of the hack I had to make for v5.3. And also > to fix vgem on non-x86. (Unfortunately changing vgem to used cached > mappings breaks x86 CI, but fixes CI on arm/arm64..) We can do that > without any changes in allocation. There is still the possibility for > problems due to cached alias, but that has been a problem this whole > time, it isn't something new.
But that just means we start exposing random low-level APIs that people will quickly abuse.. In fact even your simple plan to some extent already is an abuse of the intent of these functions, and it also requires a lot of knowledge in the driver that in the normal cases drivers can't know (e.g. is the device dma coherent or not).
