Re: [PATCH 3/3] drm/amdgpu: Try evicting from CPU visible to invisible VRAM first

[John Brooks <john-xq/Ko7C6e2Bl57MIdRCFDg@public.gmane.org>]

On Thu, May 18, 2017 at 06:08:09PM +0900, Michel Dänzer wrote:
> From: Michel Dänzer <michel.daenzer@amd.com>
> 
> In exchange, move BOs with the AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED
> flag set to CPU visible VRAM with more force.
> 
> For other BOs, this gives another chance to stay in VRAM if they
> happened to lie in the CPU visible part and another BO needs to go
> there.
> 
> This should allow BOs to stay in VRAM longer in some cases.
> 
> Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_object.c | 19 ++++++++++++++-----
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c    | 27 +++++++++++++++++++++++++++
>  2 files changed, 41 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 41ee353b22c8..d45c2325c61a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -963,11 +963,20 @@ int amdgpu_bo_fault_reserve_notify(struct ttm_buffer_object *bo)
>  					 AMDGPU_GEM_DOMAIN_GTT);
>  	abo->placements[0].lpfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
>  
> -	/* Only set GTT as busy placement; if there is no space in CPU visible
> -	 * VRAM, move this BO to GTT instead of evicting other BOs
> -	 */
> -	abo->placement.busy_placement = &abo->placements[1];
> -	abo->placement.num_busy_placement = 1;
> +	if (abo->flags & AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED) {
> +		/* Only set VRAM as normal placement; if there is no space in
> +		 * CPU visible VRAM, evict other BOs, only fall back to GTT as
> +		 * last resort
> +		 */
> +		abo->placement.num_placement = 1;
> +	} else {
> +		/* Only set GTT as busy placement; if there is no space in CPU
> +		 * visible VRAM, move this BO to GTT instead of evicting other
> +		 * BOs
> +		 */
> +		abo->placement.busy_placement = &abo->placements[1];
> +		abo->placement.num_busy_placement = 1;
> +	}
>  
>  	return ttm_bo_validate(bo, &abo->placement, false, false);
>  }
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 57789b860768..d5ed85026542 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -206,7 +206,34 @@ static void amdgpu_evict_flags(struct ttm_buffer_object *bo,
>  		    adev->mman.buffer_funcs_ring &&
>  		    adev->mman.buffer_funcs_ring->ready == false) {
>  			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_CPU);
> +		} else if (adev->mc.visible_vram_size < adev->mc.real_vram_size) {
> +			unsigned fpfn = adev->mc.visible_vram_size >> PAGE_SHIFT;
> +			struct drm_mm_node *node = bo->mem.mm_node;
> +			unsigned long pages_left;
> +
> +			for (pages_left = bo->mem.num_pages;
> +			     pages_left;
> +			     pages_left -= node->size, node++) {
> +				if (node->start < fpfn)
> +					break;
> +			}
> +
> +			if (!pages_left)
> +				goto gtt;
> +
> +			/* Try evicting to the CPU inaccessible part of VRAM
> +			 * first, but only set GTT as busy placement, so this
> +			 * BO will be evicted to GTT rather than causing other
> +			 * BOs to be evicted from VRAM
> +			 */
> +			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_VRAM |
> +							 AMDGPU_GEM_DOMAIN_GTT);
> +			abo->placements[0].fpfn = fpfn;
> +			abo->placements[0].lpfn = 0;
> +			abo->placement.busy_placement = &abo->placements[1];

Are you sure you want to hardcode the placements index? It'll be dependent on
the order set up in amdgpu_ttm_placement_init.

> +			abo->placement.num_busy_placement = 1;
>  		} else {
> +gtt:
>  			amdgpu_ttm_placement_from_domain(abo, AMDGPU_GEM_DOMAIN_GTT);
>  			/* Set an upper limit to force directly allocating
>  			 * address space for the BO.
> -- 
> 2.11.0
> 
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

--
John Brooks
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx