While this series is more correct from the DMA API point of view, it is also much heavier as it strictly disables the use of any cache on all user-space mapped BOs, and is also much more restricted in terms of which memory it can use.
I have a v4 in the works that lets us use TTM for user-mapped objects and only falls back to using the DMA API for kernel-only objects (fences & PBs basically). Will push it shortly but wanted to signal this revision as deprecated in the meantime. On Fri, Jun 27, 2014 at 8:22 PM, Alexandre Courbot <acourbot at nvidia.com> wrote: > v2 was doing some pretty nasty things with the DMA API, so I took a different > approach for this v3. > > As suggested, this version uses ttm_dma_populate() to populate BOs. The reason > for doing this was that it would entitle us to using the DMA sync functions, > but since the memory returned is already coherent anyway, we do not even > need to call these functions anymore. > > So this series has turned into 2 small patches: > > - The first attempts to make it more obvious that Nouveau can use different > ways to populate TTs, and make it possible to choose which method to use from > a single place. > - The second leverages this work to select the DMA allocator to populate TTs > on ARM. > > Doing this solves all our coherency problems with Nouveau on Tegra, and > hopefully makes the code easier to read in the process. > > Alexandre Courbot (2): > drm/nouveau: cleanup TTM population logic > drm/nouveau: use DMA TT population method on ARM > > drivers/gpu/drm/nouveau/nouveau_bo.c | 63 > ++++++++++++++++++----------------- > drivers/gpu/drm/nouveau/nouveau_drm.h | 11 ++++++ > drivers/gpu/drm/nouveau/nouveau_ttm.c | 17 ++++++++++ > 3 files changed, 61 insertions(+), 30 deletions(-) > > -- > 2.0.0 >
