[PATCH 1/2] drm/amdgpu: group BOs by log2 of the size on the LRU

[Alex Deucher]

On Thu, Apr 14, 2016 at 8:54 AM, Christian König
<deathsimple at vodafone.de> wrote:
> From: Christian König <christian.koenig at amd.com>
>
> This allows us to have small BOs on the LRU before big ones.
>
> Signed-off-by: Christian König <christian.koenig at amd.com>

Have you done any benchmarking to see how much this helps when there
is memory contention?

For the series:
Reviewed-by: Alex Deucher <alexander.deucher at amd.com>

> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu.h     | 11 ++++++
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 59 +++++++++++++++++++++++++++++++--
>  2 files changed, 68 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> index c4a21c6..7b90323 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> @@ -391,6 +391,14 @@ unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring);
>  /*
>   * TTM.
>   */
> +
> +#define AMDGPU_TTM_LRU_SIZE    20
> +
> +struct amdgpu_mman_lru {
> +       struct list_head                *lru[TTM_NUM_MEM_TYPES];
> +       struct list_head                *swap_lru;
> +};
> +
>  struct amdgpu_mman {
>         struct ttm_bo_global_ref        bo_global_ref;
>         struct drm_global_reference     mem_global_ref;
> @@ -408,6 +416,9 @@ struct amdgpu_mman {
>         struct amdgpu_ring                      *buffer_funcs_ring;
>         /* Scheduler entity for buffer moves */
>         struct amd_sched_entity                 entity;
> +
> +       /* custom LRU management */
> +       struct amdgpu_mman_lru                  log2_size[AMDGPU_TTM_LRU_SIZE];
>  };
>
>  int amdgpu_copy_buffer(struct amdgpu_ring *ring,
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index fefaa9b..b58a445 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -910,6 +910,50 @@ uint32_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
>         return flags;
>  }
>
> +static void amdgpu_ttm_lru_removal(struct ttm_buffer_object *tbo)
> +{
> +       struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
> +       unsigned i;
> +
> +       for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
> +               struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
> +
> +               if (&tbo->lru == lru->lru[tbo->mem.mem_type])
> +                       lru->lru[tbo->mem.mem_type] = tbo->lru.prev;
> +
> +               if (&tbo->swap == lru->swap_lru)
> +                       lru->swap_lru = tbo->swap.prev;
> +       }
> +}
> +
> +static struct amdgpu_mman_lru *amdgpu_ttm_lru(struct ttm_buffer_object *tbo)
> +{
> +       struct amdgpu_device *adev = amdgpu_get_adev(tbo->bdev);
> +       unsigned log2_size = min(ilog2(tbo->num_pages), AMDGPU_TTM_LRU_SIZE);
> +
> +       return &adev->mman.log2_size[log2_size];
> +}
> +
> +static struct list_head *amdgpu_ttm_lru_tail(struct ttm_buffer_object *tbo)
> +{
> +       struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
> +       struct list_head *res = lru->lru[tbo->mem.mem_type];
> +
> +       lru->lru[tbo->mem.mem_type] = &tbo->lru;
> +
> +       return res;
> +}
> +
> +static struct list_head *amdgpu_ttm_swap_lru_tail(struct ttm_buffer_object *tbo)
> +{
> +       struct amdgpu_mman_lru *lru = amdgpu_ttm_lru(tbo);
> +       struct list_head *res = lru->swap_lru;
> +
> +       lru->swap_lru = &tbo->swap;
> +
> +       return res;
> +}
> +
>  static struct ttm_bo_driver amdgpu_bo_driver = {
>         .ttm_tt_create = &amdgpu_ttm_tt_create,
>         .ttm_tt_populate = &amdgpu_ttm_tt_populate,
> @@ -923,12 +967,14 @@ static struct ttm_bo_driver amdgpu_bo_driver = {
>         .fault_reserve_notify = &amdgpu_bo_fault_reserve_notify,
>         .io_mem_reserve = &amdgpu_ttm_io_mem_reserve,
>         .io_mem_free = &amdgpu_ttm_io_mem_free,
> -       .lru_tail = &ttm_bo_default_lru_tail,
> -       .swap_lru_tail = &ttm_bo_default_swap_lru_tail,
> +       .lru_removal = &amdgpu_ttm_lru_removal,
> +       .lru_tail = &amdgpu_ttm_lru_tail,
> +       .swap_lru_tail = &amdgpu_ttm_swap_lru_tail,
>  };
>
>  int amdgpu_ttm_init(struct amdgpu_device *adev)
>  {
> +       unsigned i, j;
>         int r;
>
>         r = amdgpu_ttm_global_init(adev);
> @@ -946,6 +992,15 @@ int amdgpu_ttm_init(struct amdgpu_device *adev)
>                 DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
>                 return r;
>         }
> +
> +       for (i = 0; i < AMDGPU_TTM_LRU_SIZE; ++i) {
> +               struct amdgpu_mman_lru *lru = &adev->mman.log2_size[i];
> +
> +               for (j = 0; j < TTM_NUM_MEM_TYPES; ++j)
> +                       lru->lru[j] = &adev->mman.bdev.man[j].lru;
> +               lru->swap_lru = &adev->mman.bdev.glob->swap_lru;
> +       }
> +
>         adev->mman.initialized = true;
>         r = ttm_bo_init_mm(&adev->mman.bdev, TTM_PL_VRAM,
>                                 adev->mc.real_vram_size >> PAGE_SHIFT);
> --
> 2.5.0
>