gem/gt TODO items
-----------------
- For discrete memory manager, merge enough dg1 to be able to refactor it to
TTM. Then land pci ids (just in case that turns up an uapi problem). TTM has
improved a lot the past 2 years, there's no reason anymore not to use it.
- Come up with a plan what to do with drm/scheduler and how to get there.
- Roll out dma_fence critical section annotations.
- There's a lot of complexity added past few years to make relocations faster.
That doesn't make sense given hw and gpu apis moved away from this model years
ago:
1. Land a modern pre-bound uapi like VM_BIND
2. Any complexity added in this area past few years which can't be justified
with VM_BIND using userspace should be removed. Looking at amdgpu dma_resv on
the bo and vm, plus some lru locks is all that needed. No complex rcu,
refcounts, caching, ... on everything.
This is the matching task on the vm side compared to ttm/dma_resv on the
backing storage side.
- i915_sw_fence seems to be the main structure for the i915-gem dma_fence model.
How-to-dma_fence is core and drivers really shouldn't build their own world
here, treating everything else as a fixed platform. i915_sw_fence concepts
should be moved to dma_fence, drm/scheduler or atomic commit helpers. Or
removed if dri-devel consensus is that it's not a good idea. Once that's done
maybe even remove it if there's nothing left.
Smaller things:
- i915_utils.h needs to be moved to the right places.
- dma_fence_work should be in drivers/dma-buf
- i915_mm.c should be moved to the right places. Some of the helpers also look a
bit fishy:
https://lore.kernel.org/linux-mm/[email protected]/
- tasklet helpers in i915_tasklet.h also look a bit misplaced and should
probably be moved to tasklet headers.