[Intel-gfx] [PATCH v4 10/15] drm/ttm, drm/amdgpu: Allow the driver some control over swapping

[Christian König]

Am 26.05.21 um 13:32 schrieb Thomas Hellström:
> We are calling the eviction_valuable driver callback at eviction time to
> determine whether we actually can evict a buffer object.
> The upcoming i915 TTM backend needs the same functionality for swapout,
> and that might actually be beneficial to other drivers as well.
>
> Add an eviction_valuable call also in the swapout path. Try to keep the
> current behaviour for all drivers by returning true if the buffer object
> is already in the TTM_PL_SYSTEM placement. We change behaviour for the
> case where a buffer object is in a TT backed placement when swapped out,
> in which case the drivers normal eviction_valuable path is run.
>
> Finally make sure we don't try to swapout a bo that was recently purged
> and therefore unpopulated.
>
> Reviewed-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> Cc: Christian König <christian.koenig at amd.com>
> Signed-off-by: Thomas Hellström <thomas.hellstrom at linux.intel.com>
> ---
> v3:
> - Don't export ttm_tt_unpopulate
> - Fix confusion reading the locked pointer instead of the value
>    pointed to in ttm_bo_evict_swapout_allowable (Reported by
>    Maarten Lankhorst)
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c |  4 +++
>   drivers/gpu/drm/ttm/ttm_bo.c            | 43 ++++++++++++++++---------
>   drivers/gpu/drm/ttm/ttm_tt.c            |  3 ++
>   3 files changed, 34 insertions(+), 16 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index 3bc3aebfef7c..45d194bffc3f 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -1348,6 +1348,10 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   	struct dma_fence *f;
>   	int i;
>   
> +	/* Swapout? */
> +	if (bo->mem.mem_type == TTM_PL_SYSTEM)
> +		return true;
> +
>   	if (bo->type == ttm_bo_type_kernel &&
>   	    !amdgpu_vm_evictable(ttm_to_amdgpu_bo(bo)))
>   		return false;
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index be0406466460..1b2d062266ed 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -536,6 +536,10 @@ static int ttm_bo_evict(struct ttm_buffer_object *bo,
>   bool ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>   			      const struct ttm_place *place)
>   {
> +	dma_resv_assert_held(bo->base.resv);
> +	if (bo->mem.mem_type == TTM_PL_SYSTEM)
> +		return true;
> +
>   	/* Don't evict this BO if it's outside of the
>   	 * requested placement range
>   	 */
> @@ -558,7 +562,9 @@ EXPORT_SYMBOL(ttm_bo_eviction_valuable);
>    * b. Otherwise, trylock it.
>    */
>   static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
> -			struct ttm_operation_ctx *ctx, bool *locked, bool *busy)
> +					   struct ttm_operation_ctx *ctx,
> +					   const struct ttm_place *place,
> +					   bool *locked, bool *busy)
>   {
>   	bool ret = false;
>   
> @@ -576,6 +582,14 @@ static bool ttm_bo_evict_swapout_allowable(struct ttm_buffer_object *bo,
>   			*busy = !ret;
>   	}
>   
> +	if (ret && place && !bo->bdev->funcs->eviction_valuable(bo, place)) {
> +		ret = false;
> +		if (*locked) {
> +			dma_resv_unlock(bo->base.resv);
> +			*locked = false;
> +		}
> +	}
> +
>   	return ret;
>   }
>   
> @@ -630,20 +644,14 @@ int ttm_mem_evict_first(struct ttm_device *bdev,
>   		list_for_each_entry(bo, &man->lru[i], lru) {
>   			bool busy;
>   
> -			if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked,
> -							    &busy)) {
> +			if (!ttm_bo_evict_swapout_allowable(bo, ctx, place,
> +							    &locked, &busy)) {
>   				if (busy && !busy_bo && ticket !=
>   				    dma_resv_locking_ctx(bo->base.resv))
>   					busy_bo = bo;
>   				continue;
>   			}
>   
> -			if (place && !bdev->funcs->eviction_valuable(bo,
> -								      place)) {
> -				if (locked)
> -					dma_resv_unlock(bo->base.resv);
> -				continue;
> -			}
>   			if (!ttm_bo_get_unless_zero(bo)) {
>   				if (locked)
>   					dma_resv_unlock(bo->base.resv);
> @@ -1140,10 +1148,18 @@ EXPORT_SYMBOL(ttm_bo_wait);
>   int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
>   		   gfp_t gfp_flags)
>   {
> +	struct ttm_place place = {};
>   	bool locked;
>   	int ret;
>   
> -	if (!ttm_bo_evict_swapout_allowable(bo, ctx, &locked, NULL))
> +	/*
> +	 * While the bo may already reside in SYSTEM placement, set
> +	 * SYSTEM as new placement to cover also the move further below.
> +	 * The driver may use the fact that we're moving from SYSTEM
> +	 * as an indication that we're about to swap out.
> +	 */
> +	place.mem_type = TTM_PL_SYSTEM;
> +	if (!ttm_bo_evict_swapout_allowable(bo, ctx, &place, &locked, NULL))
>   		return -EBUSY;
>   
>   	if (!ttm_bo_get_unless_zero(bo)) {
> @@ -1168,12 +1184,7 @@ int ttm_bo_swapout(struct ttm_buffer_object *bo, struct ttm_operation_ctx *ctx,
>   	if (bo->mem.mem_type != TTM_PL_SYSTEM) {
>   		struct ttm_operation_ctx ctx = { false, false };
>   		struct ttm_resource evict_mem;
> -		struct ttm_place place, hop;
> -
> -		memset(&place, 0, sizeof(place));
> -		memset(&hop, 0, sizeof(hop));
> -
> -		place.mem_type = TTM_PL_SYSTEM;
> +		struct ttm_place hop = {};

I would stick with memset because of the padding reasons.

>   
>   		ret = ttm_resource_alloc(bo, &place, &evict_mem);
>   		if (unlikely(ret))
> diff --git a/drivers/gpu/drm/ttm/ttm_tt.c b/drivers/gpu/drm/ttm/ttm_tt.c
> index 913b330a234b..d9793cbb6d13 100644
> --- a/drivers/gpu/drm/ttm/ttm_tt.c
> +++ b/drivers/gpu/drm/ttm/ttm_tt.c
> @@ -263,6 +263,9 @@ int ttm_tt_swapout(struct ttm_device *bdev, struct ttm_tt *ttm,
>   	struct page *to_page;
>   	int i, ret;
>   
> +	if (!ttm_tt_is_populated(ttm))
> +		return 0;
> +

This here is just because of a bug in the higher level function.

I've just pushed the fix for that to drm-misc-fixes, so maybe drop that 
here as soon as this is backmerged.

Apart from that patch looks good to me.

Christian.


>   	swap_storage = shmem_file_setup("ttm swap", size, 0);
>   	if (IS_ERR(swap_storage)) {
>   		pr_err("Failed allocating swap storage\n");