Am 30.10.2017 um 04:25 schrieb Chunming Zhou:
>
>
> On 2017å¹´10æ??27æ?¥ 22:43, Christian König wrote:
>> From: Christian König <christian.koenig at amd.com>
>>
>> Just allocate the GART space and fill it.
>>
>> This prevents forcing the BO to be idle.
>>
>> v2: don't unbind/bind at all, just fill the allocated GART space
> Could you explain what 'unbind/bind' is? My old understanding is
> 'bind' is to filling gart space, isn't it?
Bind means that TTM is telling us that a ttm_tt (which represent a bunch
of system pages) is going to be used by the GPU.
Mapping the TTM to GART space is just one thing we could do during that
callback, but that is purely optional.
>
> one comment in line...
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> ---
>> Â drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 21 +++++++++++++++------
>> Â 1 file changed, 15 insertions(+), 6 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> index eead348..06871a9 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> @@ -878,10 +878,12 @@ static int amdgpu_ttm_backend_bind(struct
>> ttm_tt *ttm,
>> Â int amdgpu_ttm_alloc_gart(struct ttm_buffer_object *bo)
>> Â {
>> Â Â Â Â Â struct amdgpu_device *adev = amdgpu_ttm_adev(bo->bdev);
>> +Â Â Â struct amdgpu_ttm_tt *gtt = (void*)bo->ttm;
>> Â Â Â Â Â struct ttm_mem_reg tmp;
>> Â Â Â Â Â Â struct ttm_placement placement;
>> Â Â Â Â Â struct ttm_place placements;
>> +Â Â Â uint64_t flags;
>> Â Â Â Â Â int r;
>> Â Â Â Â Â Â if (bo->mem.mem_type != TTM_PL_TT ||
>> @@ -903,14 +905,21 @@ int amdgpu_ttm_alloc_gart(struct
>> ttm_buffer_object *bo)
>> Â Â Â Â Â if (unlikely(r))
>> Â Â Â Â Â Â Â Â Â return r;
>> Â -Â Â Â r = ttm_bo_move_ttm(bo, true, false, &tmp);
>> -Â Â Â if (unlikely(r))
>> +Â Â Â flags = amdgpu_ttm_tt_pte_flags(adev, bo->ttm, &tmp);
>> +Â Â Â gtt->offset = (u64)tmp.start << PAGE_SHIFT;
>> +Â Â Â r = amdgpu_gart_bind(adev, gtt->offset, bo->ttm->num_pages,
>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â bo->ttm->pages, gtt->ttm.dma_address, flags);
>> +Â Â Â if (unlikely(r)) {
>> Â Â Â Â Â Â Â Â Â ttm_bo_mem_put(bo, &tmp);
>> -Â Â Â else
>> -Â Â Â Â Â Â Â bo->offset = (bo->mem.start << PAGE_SHIFT) +
>> - bo->bdev->man[bo->mem.mem_type].gpu_offset;
>> +Â Â Â Â Â Â Â return r;
>> +Â Â Â }
>> Â -Â Â Â return r;
>> +Â Â Â ttm_bo_mem_put(bo, &bo->mem);
> The old node doesn't need to wait before put it? if another user
> allocate in same place immediately, doesn't it result in problem?
Good point, but no that isn't a problem.
When we go into this branch we know that the old node didn't had any
GART space allocated to it, so we don't need to wait for anything using
this GART space.
Regards,
Christian.
>
> Regards,
> David Zhou
>> +Â Â Â bo->mem = tmp;
>> +Â Â Â bo->offset = (bo->mem.start << PAGE_SHIFT) +
>> +Â Â Â Â Â Â Â bo->bdev->man[bo->mem.mem_type].gpu_offset;
>> +
>> +Â Â Â return 0;
>> Â }
>> Â Â int amdgpu_ttm_recover_gart(struct ttm_buffer_object *tbo)
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
