Am 02.12.21 um 05:55 schrieb xinhui pan:
TTM want bo->resource to be valid during BO's life.
But ttm_bo_mem_space might fail and bo->resource point to NULL. Many code
touch bo->resource and hit panic then.
As old and new mem might overlap, move ttm_resource_free after
ttm_bo_mem_space is not an option.
We could assign one sysmem node to BO to make bo->resource valid.
Signed-off-by: xinhui pan <xinhui.pan@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 14 ++++++++++++--
1 file changed, 12 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
index c4317343967f..697fac0b82a3 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
@@ -359,6 +359,7 @@ int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
struct amdgpu_bo **bo_ptr, void **cpu_addr)
{
struct ttm_operation_ctx ctx = { false, false };
+ struct ttm_resource *tmp_res;
unsigned int i;
int r;
@@ -380,17 +381,26 @@ int amdgpu_bo_create_kernel_at(struct amdgpu_device *adev,
if (cpu_addr)
amdgpu_bo_kunmap(*bo_ptr);
- ttm_resource_free(&(*bo_ptr)->tbo, &(*bo_ptr)->tbo.resource);
+ /* Assign one sysmem node to BO as we want bo->resource to be valid. */
+ amdgpu_bo_placement_from_domain(*bo_ptr, AMDGPU_GEM_DOMAIN_CPU);
+ r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
+ &tmp_res, &ctx);
+ if (r)
+ goto error;
+
+ ttm_bo_move_null(&(*bo_ptr)->tbo, tmp_res);
That's much easier fixable if you just create the BO in the CPU domain
in the first place.
E.g. modify the call to amdgpu_bo_create_reserved() above to use
AMDGPU_GEM_DOMAIN_CPU instead of the provided one.
This way you can indeed move the ttm_resource_free() after the assignment.
Regards,
Christian.
for (i = 0; i < (*bo_ptr)->placement.num_placement; ++i) {
(*bo_ptr)->placements[i].fpfn = offset >> PAGE_SHIFT;
(*bo_ptr)->placements[i].lpfn = (offset + size) >> PAGE_SHIFT;
}
r = ttm_bo_mem_space(&(*bo_ptr)->tbo, &(*bo_ptr)->placement,
- &(*bo_ptr)->tbo.resource, &ctx);
+ &tmp_res, &ctx);
if (r)
goto error;
+ ttm_bo_move_null(&(*bo_ptr)->tbo, tmp_res);
+
if (cpu_addr) {
r = amdgpu_bo_kmap(*bo_ptr, cpu_addr);
if (r)
