Re: [PATCH v4 1/3] drm/amdgpu: UVD avoid memory allocation during IB test

["Christian König" <christian.koenig@amd.com>]

Am 13.09.21 um 10:42 schrieb xinhui pan:
> move BO allocation in sw_init.
>
> Signed-off-by: xinhui pan <xinhui.pan@amd.com>
> ---
> change from v3:
> drop the bo resv lock in ib test.
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c | 102 ++++++++++++++++--------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.h |   1 +
>   drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c   |  11 +--
>   drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c   |  11 +--
>   4 files changed, 72 insertions(+), 53 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> index d451c359606a..b0fbd5a1d5af 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.c
> @@ -134,6 +134,51 @@ MODULE_FIRMWARE(FIRMWARE_VEGA12);
>   MODULE_FIRMWARE(FIRMWARE_VEGA20);
>   
>   static void amdgpu_uvd_idle_work_handler(struct work_struct *work);
> +static void amdgpu_uvd_force_into_uvd_segment(struct amdgpu_bo *abo);
> +
> +static int amdgpu_uvd_create_msg_bo_helper(struct amdgpu_device *adev,
> +					   uint32_t size,
> +					   struct amdgpu_bo **bo_ptr)
> +{
> +	struct ttm_operation_ctx ctx = { true, false };
> +	struct amdgpu_bo *bo = NULL;
> +	void *addr;
> +	int r;
> +
> +	r = amdgpu_bo_create_reserved(adev, size, PAGE_SIZE,
> +				      AMDGPU_GEM_DOMAIN_GTT,
> +				      &bo, NULL, &addr);
> +	if (r)
> +		return r;
> +
> +	if (adev->uvd.address_64_bit)
> +		goto succ;
> +
> +	amdgpu_bo_kunmap(bo);
> +	amdgpu_bo_unpin(bo);
> +	amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
> +	amdgpu_uvd_force_into_uvd_segment(bo);
> +	r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
> +	if (r)
> +		goto err;
> +	r = amdgpu_bo_pin(bo, AMDGPU_GEM_DOMAIN_VRAM);
> +	if (r)
> +		goto err_pin;
> +	r = amdgpu_bo_kmap(bo, &addr);
> +	if (r)
> +		goto err_kmap;
> +succ:
> +	amdgpu_bo_unreserve(bo);
> +	*bo_ptr = bo;
> +	return 0;
> +err_kmap:
> +	amdgpu_bo_unpin(bo);
> +err_pin:
> +err:
> +	amdgpu_bo_unreserve(bo);
> +	amdgpu_bo_unref(&bo);
> +	return r;
> +}
>   
>   int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
>   {
> @@ -302,6 +347,10 @@ int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
>   	if (!amdgpu_device_ip_block_version_cmp(adev, AMD_IP_BLOCK_TYPE_UVD, 5, 0))
>   		adev->uvd.address_64_bit = true;
>   
> +	r = amdgpu_uvd_create_msg_bo_helper(adev, 128 << 10, &adev->uvd.ib_bo);
> +	if (r)
> +		return r;
> +
>   	switch (adev->asic_type) {
>   	case CHIP_TONGA:
>   		adev->uvd.use_ctx_buf = adev->uvd.fw_version >= FW_1_65_10;
> @@ -324,6 +373,7 @@ int amdgpu_uvd_sw_init(struct amdgpu_device *adev)
>   
>   int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
>   {
> +	void *addr = amdgpu_bo_kptr(adev->uvd.ib_bo);
>   	int i, j;
>   
>   	drm_sched_entity_destroy(&adev->uvd.entity);
> @@ -342,6 +392,7 @@ int amdgpu_uvd_sw_fini(struct amdgpu_device *adev)
>   		for (i = 0; i < AMDGPU_MAX_UVD_ENC_RINGS; ++i)
>   			amdgpu_ring_fini(&adev->uvd.inst[j].ring_enc[i]);
>   	}
> +	amdgpu_bo_free_kernel(&adev->uvd.ib_bo, NULL, &addr);
>   	release_firmware(adev->uvd.fw);
>   
>   	return 0;
> @@ -1080,23 +1131,10 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
>   	unsigned offset_idx = 0;
>   	unsigned offset[3] = { UVD_BASE_SI, 0, 0 };
>   
> -	amdgpu_bo_kunmap(bo);
> -	amdgpu_bo_unpin(bo);
> -
> -	if (!ring->adev->uvd.address_64_bit) {
> -		struct ttm_operation_ctx ctx = { true, false };
> -
> -		amdgpu_bo_placement_from_domain(bo, AMDGPU_GEM_DOMAIN_VRAM);
> -		amdgpu_uvd_force_into_uvd_segment(bo);
> -		r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
> -		if (r)
> -			goto err;
> -	}
> -
>   	r = amdgpu_job_alloc_with_ib(adev, 64, direct ? AMDGPU_IB_POOL_DIRECT :
>   				     AMDGPU_IB_POOL_DELAYED, &job);
>   	if (r)
> -		goto err;
> +		return r;
>   
>   	if (adev->asic_type >= CHIP_VEGA10) {
>   		offset_idx = 1 + ring->me;
> @@ -1148,8 +1186,6 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
>   	}
>   
>   	amdgpu_bo_fence(bo, f, false);
> -	amdgpu_bo_unreserve(bo);
> -	amdgpu_bo_unref(&bo);
>   
>   	if (fence)
>   		*fence = dma_fence_get(f);
> @@ -1159,10 +1195,6 @@ static int amdgpu_uvd_send_msg(struct amdgpu_ring *ring, struct amdgpu_bo *bo,
>   
>   err_free:
>   	amdgpu_job_free(job);
> -
> -err:
> -	amdgpu_bo_unreserve(bo);
> -	amdgpu_bo_unref(&bo);
>   	return r;
>   }
>   
> @@ -1173,16 +1205,11 @@ int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
>   			      struct dma_fence **fence)
>   {
>   	struct amdgpu_device *adev = ring->adev;
> -	struct amdgpu_bo *bo = NULL;
> +	struct amdgpu_bo *bo = adev->uvd.ib_bo;
>   	uint32_t *msg;
>   	int r, i;
>   
> -	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_GTT,
> -				      &bo, NULL, (void **)&msg);
> -	if (r)
> -		return r;
> -
> +	msg = amdgpu_bo_kptr(bo);
>   	/* stitch together an UVD create msg */
>   	msg[0] = cpu_to_le32(0x00000de4);
>   	msg[1] = cpu_to_le32(0x00000000);
> @@ -1198,7 +1225,9 @@ int amdgpu_uvd_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
>   	for (i = 11; i < 1024; ++i)
>   		msg[i] = cpu_to_le32(0x0);
>   
> -	return amdgpu_uvd_send_msg(ring, bo, true, fence);
> +	r = amdgpu_uvd_send_msg(ring, bo, true, fence);
> +
> +	return r;

Drop that change, the code can now keep as it is.

>   }
>   
>   int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
> @@ -1207,14 +1236,16 @@ int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
>   	struct amdgpu_device *adev = ring->adev;
>   	struct amdgpu_bo *bo = NULL;
>   	uint32_t *msg;
> -	int r, i;
> +	int r = 0, i;
>   
> -	r = amdgpu_bo_create_reserved(adev, 1024, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_GTT,
> -				      &bo, NULL, (void **)&msg);
> +	if (direct)
> +		bo = adev->uvd.ib_bo;
> +	else
> +		r = amdgpu_uvd_create_msg_bo_helper(adev, 4096, &bo);
>   	if (r)
>   		return r;

Better use { r= ..; if(r) return r; } instead of initializing r to zero 
above. Initializing r like that is often seen as bad style.

Apart from those two nit picks the patch is Reviewed-by: Christian König 
<christian.koenig@amd.com>

Thanks,
Christian.

>   
> +	msg = amdgpu_bo_kptr(bo);
>   	/* stitch together an UVD destroy msg */
>   	msg[0] = cpu_to_le32(0x00000de4);
>   	msg[1] = cpu_to_le32(0x00000002);
> @@ -1223,7 +1254,12 @@ int amdgpu_uvd_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
>   	for (i = 4; i < 1024; ++i)
>   		msg[i] = cpu_to_le32(0x0);
>   
> -	return amdgpu_uvd_send_msg(ring, bo, direct, fence);
> +	r = amdgpu_uvd_send_msg(ring, bo, direct, fence);
> +
> +	if (!direct)
> +		amdgpu_bo_free_kernel(&bo, NULL, (void **)&msg);
> +
> +	return r;
>   }
>   
>   static void amdgpu_uvd_idle_work_handler(struct work_struct *work)
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.h
> index edbb8194ee81..76ac9699885d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_uvd.h
> @@ -68,6 +68,7 @@ struct amdgpu_uvd {
>   	/* store image width to adjust nb memory state */
>   	unsigned		decode_image_width;
>   	uint32_t                keyselect;
> +	struct amdgpu_bo	*ib_bo;
>   };
>   
>   int amdgpu_uvd_sw_init(struct amdgpu_device *adev);
> diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c
> index bc571833632e..d5d023a24269 100644
> --- a/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v6_0.c
> @@ -332,15 +332,9 @@ static int uvd_v6_0_enc_get_destroy_msg(struct amdgpu_ring *ring,
>   static int uvd_v6_0_enc_ring_test_ib(struct amdgpu_ring *ring, long timeout)
>   {
>   	struct dma_fence *fence = NULL;
> -	struct amdgpu_bo *bo = NULL;
> +	struct amdgpu_bo *bo = ring->adev->uvd.ib_bo;
>   	long r;
>   
> -	r = amdgpu_bo_create_reserved(ring->adev, 128 * 1024, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_VRAM,
> -				      &bo, NULL, NULL);
> -	if (r)
> -		return r;
> -
>   	r = uvd_v6_0_enc_get_create_msg(ring, 1, bo, NULL);
>   	if (r)
>   		goto error;
> @@ -357,9 +351,6 @@ static int uvd_v6_0_enc_ring_test_ib(struct amdgpu_ring *ring, long timeout)
>   
>   error:
>   	dma_fence_put(fence);
> -	amdgpu_bo_unpin(bo);
> -	amdgpu_bo_unreserve(bo);
> -	amdgpu_bo_unref(&bo);
>   	return r;
>   }
>   
> diff --git a/drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c b/drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c
> index b6e82d75561f..c115b2da22ef 100644
> --- a/drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/uvd_v7_0.c
> @@ -338,15 +338,9 @@ static int uvd_v7_0_enc_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handl
>   static int uvd_v7_0_enc_ring_test_ib(struct amdgpu_ring *ring, long timeout)
>   {
>   	struct dma_fence *fence = NULL;
> -	struct amdgpu_bo *bo = NULL;
> +	struct amdgpu_bo *bo = ring->adev->uvd.ib_bo;
>   	long r;
>   
> -	r = amdgpu_bo_create_reserved(ring->adev, 128 * 1024, PAGE_SIZE,
> -				      AMDGPU_GEM_DOMAIN_VRAM,
> -				      &bo, NULL, NULL);
> -	if (r)
> -		return r;
> -
>   	r = uvd_v7_0_enc_get_create_msg(ring, 1, bo, NULL);
>   	if (r)
>   		goto error;
> @@ -363,9 +357,6 @@ static int uvd_v7_0_enc_ring_test_ib(struct amdgpu_ring *ring, long timeout)
>   
>   error:
>   	dma_fence_put(fence);
> -	amdgpu_bo_unpin(bo);
> -	amdgpu_bo_unreserve(bo);
> -	amdgpu_bo_unref(&bo);
>   	return r;
>   }
>