[PATCH] drm/amdkfd: reserve a fence slot while locking the BO

[Felix Kuehling]

On 2023-10-20 08:33, Christian König wrote:
> Looks like the KFD still needs this.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>
> Fixes: 8abc1eb2987a ("drm/amdkfd: switch over to using drm_exec v3")

To fix the immediate problem, this patch is

Acked-by: Felix Kuehling <Felix.Kuehling at amd.com>

As I understand it, this reserves a fence slot for adding an eviction 
fence. I'm not convinced that this is the right place to do this. Not 
all callers of reserve_bo_and_vm add eviction fences. In another patch 
series I added the fence slot reservation in a new helper function 
amdgpu_amdkfd_bo_validate_and_fence.

Taking another step back, as I understand it, the pre-reservation of 
fence slots is there to avoid late failures after submitting commands to 
the HW. This isn't really a problem for KFD because eviction fences 
aren't directly linked to commands submitted to the HW. It's more like a 
place holder for future user mode submissions. So I think it's OK to 
reserve the fence slot just before attaching the fence to a BO resv. We 
don't have to do the pre-reservation here.

Regards,
   Felix


> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> index 7d6daf8d2bfa..e036011137aa 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c
> @@ -1103,7 +1103,7 @@ static int reserve_bo_and_vm(struct kgd_mem *mem,
>   		if (unlikely(ret))
>   			goto error;
>   
> -		ret = drm_exec_lock_obj(&ctx->exec, &bo->tbo.base);
> +		ret = drm_exec_prepare_obj(&ctx->exec, &bo->tbo.base, 1);
>   		drm_exec_retry_on_contention(&ctx->exec);
>   		if (unlikely(ret))
>   			goto error;