[PATCH 5/5] drm/amdgpu/uvd: Move BOs to GTT when we can't place them in VRAM correctly

["Timur Kristóf" <timur.kristof@gmail.com>]

When VRAM is nearly full, the Buddy allocator makes tradeoffs
and it may place BOs in a way that they cross 256M segments.

Move the BO to GTT when this eventuality is detected.

Closes: https://gitlab.freedesktop.org/drm/amd/-/work_items/4800
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c |  3 ++-
 drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 21 +++++++++++++++++++++
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index a106c7e77e26..fb49bd53bd00 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -977,6 +977,7 @@ u32 amdgpu_ttm_fill_gart_256M_placements(struct ttm_buffer_object *bo,
 					 u32 max_placements)
 {
 	struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
+	const u64 sz = adev->gmc.gart_size;
 	u32 i;
 
 	/* Fill the placements array with 256M segments, starting from highest. */
@@ -984,7 +985,7 @@ u32 amdgpu_ttm_fill_gart_256M_placements(struct ttm_buffer_object *bo,
 		if (i * SZ_256M >= adev->gmc.gart_size)
 			break;
 
-		placements[i].lpfn = (adev->gmc.gart_size - i * SZ_256M) >> PAGE_SHIFT;
+		placements[i].lpfn = MIN(ALIGN(sz, SZ_256M) - i * SZ_256M, sz) >> PAGE_SHIFT;
 		placements[i].fpfn = ALIGN_DOWN(placements[i].lpfn - 1, SZ_256M >> PAGE_SHIFT);
 		placements[i].mem_type = TTM_PL_TT;
 		placements[i].flags = bo->resource->placement;
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
index 993957927782..53f810c2a5fb 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
@@ -617,6 +617,27 @@ static int amdgpu_uvd_cs_pass1(struct amdgpu_uvd_cs_ctx *ctx)
 			amdgpu_uvd_force_into_uvd_segment(bo);
 
 		r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
+		if (r)
+			return r;
+
+		/* Check if the BO placement crosses a 256M segment. */
+		if ((amdgpu_bo_gpu_offset(bo) >> 28) !=
+		    ((amdgpu_bo_gpu_offset(bo) + amdgpu_bo_size(bo)) >> 28)) {
+			/* There is not enough memory for correct placement of FB/MSG BOs. */
+			if (cmd == 0x0 || cmd == 0x3)
+				return -ENOMEM;
+
+			/* GTT->GTT moves are not implemented yet. */
+			if (bo->tbo.resource->mem_type != TTM_PL_VRAM)
+				return -ENOMEM;
+
+			/* Try to move the BO from VRAM to GART into a 256M segment. */
+			amdgpu_ttm_fill_gart_256M_placements(&bo->tbo,
+							     bo->placements,
+							     ARRAY_SIZE(bo->placements));
+
+			r = ttm_bo_validate(&bo->tbo, &bo->placement, &tctx);
+		}
 	}
 
 	return r;
-- 
2.54.0