On 2021-11-19 5:17 a.m., xinhui pan wrote:
amdgpu_amdkfd_gpuvm_free_memory_of_gpu drop dmabuf reference increased in amdgpu_gem_prime_export. amdgpu_bo_destroy drop dmabuf reference increased in amdgpu_gem_prime_import.
OK. This was not obvious at all. The get_dma_buf in amdgpu_gem_prime_import could use a comment that it's balanced by the dma_buf_put in drm_prime_gem_destroy, which gets called from amdgpu_bo_destroy for any bo that has a bo->tbo.base.import_attach.
I confirmed the bug and tested the fix by running the following test on a 2-GPU system
HSA_USERPTR_FOR_PAGED_MEM=0 run_kfdtest.sh --gtest_filter=KFDMemoryTest.MMBench
So remove this extra dma_buf_put to avoid double free. Signed-off-by: xinhui pan <[email protected]>
Tested-by: Felix Kuehling <[email protected]> Reviewed-by: Felix Kuehling <[email protected]>
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ------ 1 file changed, 6 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c index 6c6d5d994b5d..0f68d28e002e 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c @@ -665,12 +665,6 @@ kfd_mem_attach_dmabuf(struct amdgpu_device *adev, struct kgd_mem *mem, if (IS_ERR(gobj)) return PTR_ERR(gobj);- /* Import takes an extra reference on the dmabuf. Drop it now to- * avoid leaking it. We only need the one reference in - * kgd_mem->dmabuf. - */ - dma_buf_put(mem->dmabuf); - *bo = gem_to_amdgpu_bo(gobj); (*bo)->flags |= AMDGPU_GEM_CREATE_PREEMPTIBLE; (*bo)->parent = amdgpu_bo_ref(mem->bo);
