[PATCH 0/2] drm/gpuvm: avoid deferred cleanup on GpuVmBoAlloc drop

[Daniel Pesic]

GpuVmBoAlloc::drop() (rust/kernel/drm/gpuvm/vm_bo.rs) currently releases
its drm_gpuvm_bo via drm_gpuvm_bo_put_deferred(), which exists to handle
destruction from contexts where the GEM's gpuva lock cannot be taken
synchronously. However, a GpuVmBoAlloc has a refcount of one and is
never added to the gem, extobj, or evict lists by construction.
Therefore, a GpuVmBoAlloc that is dropped without being consumed by
obtain() does not require a deferred put.

drm_gpuvm_bo_obtain_prealloc() currently handles this situation
internally, where it discards the losing preallocated drm_gpuvm_bo by
unlocking obj->gpuva.lock and calling
drm_gpuvm_bo_destroy_not_in_lists() directly, which skips the deferred
path.

This series lets GpuVmBoAlloc::drop() behave the same:

	1/2: exports drm_gpuvm_bo_destroy_not_in_lists() so it can be
	called from the Rust binding, and promotes its comment to
	kernel-doc.

	2/2: switches GpuVmBoAlloc::drop() to call it directly, and
	includes the safety argument for why its precondition holds.
	Drops the TODO left previously.

Note: there is currently no in-tree caller of GpuVm::obtain(), so the
drop path isn't currently exercised by an existing driver.

Daniel Pesic (2):
  drm/gpuvm: export drm_gpuvm_bo_destroy_not_in_lists()
  rust: drm: gpuvm: call drm_gpuvm_bo_destroy_not_in_lists() in Drop

 drivers/gpu/drm/drm_gpuvm.c    |  5 +++--
 include/drm/drm_gpuvm.h        |  2 ++
 rust/kernel/drm/gpuvm/vm_bo.rs | 10 +++++++---
 3 files changed, 12 insertions(+), 5 deletions(-)

-- 
2.55.0