In the interest of sharing as much of possible of the userptr design, two rfc patches that are aiming to reduce the userptr notifier_lock contention.
We have these changes already in i915, but we wanted to lift them to dri- devel in the amdgpu context for discussion to make sure we don't miss anything important.
The first patch addresses the case where idling a single userptr bo for even a simple invalidation, stalls command submission across the device. Since the notifier callback is guaranteed to run to *completion* by the notifier code without a racing successful command submission, moving the fence wait out of the notifier lock should be safe.
The second patch probably has a smaller impact and might avoid some contention particularly if traversing long userptr lists. The notifier sem is also locked interruptible during command submission.
Any comments appreciated.
Cc: Christian König christian.koenig@amd.com
Thomas Hellström (2): drm/amdgpu: Move dma_resv waiting outside the notifier lock drm/amdgpu: Use an rwsem for the notifier lock
drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 12 ++++++++---- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 2 +- drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c | 11 ++++++----- drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 2 +- 6 files changed, 18 insertions(+), 13 deletions(-)