[PATCH 3/4] drm/amdgpu: Make amdgpu_vram_mgr_bo_invisible_size always accurate

[Christian König]

Am 15.06.2018 um 17:14 schrieb Michel Dänzer:
> On 2018-06-15 05:11 PM, Christian König wrote:
>> Am 15.06.2018 um 16:42 schrieb Michel Dänzer:
>>> From: Michel Dänzer <michel.daenzer at amd.com>
>>>
>>> Even BOs with AMDGPU_GEM_CREATE_NO_CPU_ACCESS may end up at least
>>> partially in CPU visible VRAM, in particular when all VRAM is visible.
>>>
>>> Cc: stable at vger.kernel.org
>>> Signed-off-by: Michel Dänzer <michel.daenzer at amd.com>
>>> ---
>>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 26 ++++++++++++++++++--
>>>    1 file changed, 24 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> index 491fd0a789f7..9b34df7017f0 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>>> @@ -106,10 +106,32 @@ static u64 amdgpu_vram_mgr_vis_size(struct
>>> amdgpu_device *adev,
>>>     */
>>>    u64 amdgpu_vram_mgr_bo_invisible_size(struct amdgpu_bo *bo)
>>>    {
>>> -    if (bo->flags & AMDGPU_GEM_CREATE_NO_CPU_ACCESS)
>>> +    struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev);
>>> +    struct amdgpu_vram_mgr *mgr = adev->mman.bdev.man[TTM_PL_VRAM].priv;
>>> +    struct ttm_mem_reg *mem = &bo->tbo.mem;
>>> +    struct drm_mm_node *nodes = mem->mm_node;
>>> +    unsigned pages = mem->num_pages;
>>> +    u64 usage = 0;
>>> +
>>> +    if (adev->gmc.visible_vram_size == adev->gmc.real_vram_size)
>>> +        return 0;
>> Andrey just added a helper for that into amdgpu_gmc.h.
> Yep, see patch 4 and the cover letter. :)

Yeah, seen that after hitting send :)

>>> +
>>> +    if (mem->start >= adev->gmc.visible_vram_size >> PAGE_SHIFT)
>>>            return amdgpu_bo_size(bo);
>>>    -    return 0;
>>> +    if (!nodes)
>>> +        return 0;
>>> +
>>> +    spin_lock(&mgr->lock);
>> I actually don't see any need to grab the lock here.
> Ah, so amdgpu_vram_mgr_del only grabs it for drm_mm_remove_node?

Yes, exactly.

Christian.