[PATCH v4 2/3] drm/amdgpu: VCE avoid memory allocation during IB test

[Liu, Leo]

[AMD Official Use Only]

256 bytes alignment is for Video HW that is with GFX9, so it should be fine in general.

Regards,
Leo


-----Original Message-----
From: Koenig, Christian <Christian.Koenig at amd.com>
Sent: September 13, 2021 5:04 AM
To: Pan, Xinhui <Xinhui.Pan at amd.com>; amd-gfx at lists.freedesktop.org
Cc: Deucher, Alexander <Alexander.Deucher at amd.com>; Liu, Leo <Leo.Liu at amd.com>; Zhu, James <James.Zhu at amd.com>
Subject: Re: [PATCH v4 2/3] drm/amdgpu: VCE avoid memory allocation during IB test

Am 13.09.21 um 10:42 schrieb xinhui pan:
> alloc extra msg from direct IB pool.
>
> Signed-off-by: xinhui pan <xinhui.pan at amd.com>

It would be cleaner if Leo could confirm that 256 byte alignment would work as well.

But either way Reviewed-by: Christian König <christian.koenig at amd.com>

Regards,
Christian.

> ---
> change from v1:
> msg is allocated separately.
> msg is aligned to gpu page boundary
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c | 27 ++++++++++++-------------
>   1 file changed, 13 insertions(+), 14 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
> index e9fdf49d69e8..caa4d3420e00 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vce.c
> @@ -82,7 +82,6 @@ MODULE_FIRMWARE(FIRMWARE_VEGA20);
>
>   static void amdgpu_vce_idle_work_handler(struct work_struct *work);
>   static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
> -                                  struct amdgpu_bo *bo,
>                                    struct dma_fence **fence);
>   static int amdgpu_vce_get_destroy_msg(struct amdgpu_ring *ring, uint32_t handle,
>                                     bool direct, struct dma_fence **fence); @@ -441,12 +440,12
> @@ void amdgpu_vce_free_handles(struct amdgpu_device *adev, struct drm_file *filp)
>    * Open up a stream for HW test
>    */
>   static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
> -                                  struct amdgpu_bo *bo,
>                                    struct dma_fence **fence)
>   {
>       const unsigned ib_size_dw = 1024;
>       struct amdgpu_job *job;
>       struct amdgpu_ib *ib;
> +     struct amdgpu_ib ib_msg;
>       struct dma_fence *f = NULL;
>       uint64_t addr;
>       int i, r;
> @@ -456,9 +455,17 @@ static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
>       if (r)
>               return r;
>
> -     ib = &job->ibs[0];
> +     memset(&ib_msg, 0, sizeof(ib_msg));
> +     /* only one gpu page is needed, alloc +1 page to make addr aligned. */
> +     r = amdgpu_ib_get(ring->adev, NULL, AMDGPU_GPU_PAGE_SIZE * 2,
> +                       AMDGPU_IB_POOL_DIRECT,
> +                       &ib_msg);
> +     if (r)
> +             goto err;
>
> -     addr = amdgpu_bo_gpu_offset(bo);
> +     ib = &job->ibs[0];
> +     /* let addr point to page boundary */
> +     addr = AMDGPU_GPU_PAGE_ALIGN(ib_msg.gpu_addr);
>
>       /* stitch together an VCE create msg */
>       ib->length_dw = 0;
> @@ -498,6 +505,7 @@ static int amdgpu_vce_get_create_msg(struct amdgpu_ring *ring, uint32_t handle,
>               ib->ptr[i] = 0x0;
>
>       r = amdgpu_job_submit_direct(job, ring, &f);
> +     amdgpu_ib_free(ring->adev, &ib_msg, f);
>       if (r)
>               goto err;
>
> @@ -1134,20 +1142,13 @@ int amdgpu_vce_ring_test_ring(struct amdgpu_ring *ring)
>   int amdgpu_vce_ring_test_ib(struct amdgpu_ring *ring, long timeout)
>   {
>       struct dma_fence *fence = NULL;
> -     struct amdgpu_bo *bo = NULL;
>       long r;
>
>       /* skip vce ring1/2 ib test for now, since it's not reliable */
>       if (ring != &ring->adev->vce.ring[0])
>               return 0;
>
> -     r = amdgpu_bo_create_reserved(ring->adev, 512, PAGE_SIZE,
> -                                   AMDGPU_GEM_DOMAIN_VRAM,
> -                                   &bo, NULL, NULL);
> -     if (r)
> -             return r;
> -
> -     r = amdgpu_vce_get_create_msg(ring, 1, bo, NULL);
> +     r = amdgpu_vce_get_create_msg(ring, 1, NULL);
>       if (r)
>               goto error;
>
> @@ -1163,8 +1164,6 @@ int amdgpu_vce_ring_test_ib(struct amdgpu_ring
> *ring, long timeout)
>
>   error:
>       dma_fence_put(fence);
> -     amdgpu_bo_unreserve(bo);
> -     amdgpu_bo_free_kernel(&bo, NULL, NULL);
>       return r;
>   }
>