[PATCH 1/2] drm/amdgpu: re-validate per VM BOs if required

[Christian König]

Am 26.03.2018 um 10:47 schrieb Chunming Zhou:
> From: Christian König <ckoenig.leichtzumerken at gmail.com>
>
> If a per VM BO ends up in a allowed domain it never moves back into the
> prefered domain.
>
> Change-Id: Ifb3e561785d3b464da28c439b271c26825224c5e
> Signed-off-by: Christian König <christian.koenig at amd.com>
> Reviewed-and-Tested-by: Chunming Zhou <david1.zhou at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 15 +++++++++++++--
>   1 file changed, 13 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index e9a41dd05345..7db411a282ce 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -1808,14 +1808,16 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
>   
>   	spin_lock(&vm->status_lock);
>   	while (!list_empty(&vm->moved)) {
> -		struct amdgpu_bo_va *bo_va;
>   		struct reservation_object *resv;
> +		struct amdgpu_bo_va *bo_va;
> +		struct amdgpu_bo *bo;
>   
>   		bo_va = list_first_entry(&vm->moved,
>   			struct amdgpu_bo_va, base.vm_status);
>   		spin_unlock(&vm->status_lock);
>   
> -		resv = bo_va->base.bo->tbo.resv;
> +		bo = bo_va->base.bo;
> +		resv = bo->tbo.resv;
>   
>   		/* Per VM BOs never need to bo cleared in the page tables */
>   		if (resv == vm->root.base.bo->tbo.resv)
> @@ -1835,6 +1837,15 @@ int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
>   			kcl_reservation_object_unlock(resv);
>   
>   		spin_lock(&vm->status_lock);
> +
> +		/* If the BO prefers to be in VRAM, but currently isn't add it
> +		 * back to the evicted list so that it gets validated again on
> +		 * the next command submission.
> +		 */
> +		if (resv == vm->root.base.bo->tbo.resv &&
> +		    bo->preferred_domains == AMDGPU_GEM_DOMAIN_VRAM &&
> +		    bo->tbo.mem.mem_type != TTM_PL_VRAM)
> +			list_add_tail(&bo_va->base.vm_status, &vm->evicted);

Thinking about this more, might be even better to put that into 
amdgpu_vm_bo_update() instead of the 
"list_del_init(&bo_va->base.vm_status);".

I'm on easter vacation this week, so checking mails is delayed. But I'm 
going to look at patch #2 this evening or tomorrow.

Regards,
Christian.

>   	}
>   	spin_unlock(&vm->status_lock);
>