Re: 答复: 答复: [PATCH 1/2] drm/amdgpu: disable the VRAM manager on special placements v2

[Christian König]

Yeah, please send that out.

Maybe just to me first, since we probably need to simplify and clean it 
up a bit, e.g. use a callback instead of duplicating much code etc...

But in general that sounded like a step into the right direction in out 
internal discussion.

We just need to change the placement as well to avoid this stupid moving 
around the buffer on every command submission.

Regards,
Christian.

Am 07.11.2016 um 09:16 schrieb Qu, Jim:
> Hi Christian:
>
> Could the patch for mmap splited pages be applied to ttm? if so, I can sent out the patch.
>
> Thanks
> JimQu
>
> ________________________________________
> 发件人: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> 代表 Christian König <deathsimple at vodafone.de>
> 发送时间: 2016年11月7日 16:12
> 收件人: Qu, Jim; amd-gfx at lists.freedesktop.org
> 主题: Re: 答复: [PATCH 1/2] drm/amdgpu: disable the VRAM manager on special placements v2
>
>> Here , checking place->lpfn , is that mean if required place is located in visible vram, driver run none pages split code path?
> Yes, exactly. CPU accessible allocations weren't meant to be split with
> the initial implementation.
>
> For this we need to be able map scattered allocations into the CPU VM
> page tables first.
>
> You kind of implemented this, but it would still cause a ping/pong of
> the BO inside the visible VRAM space.
>
> We would need to move the handling of
> AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED from amdgpu_ttm_placement_init()
> down into the VRAM manager as well.
>
> Regards,
> Christian.
>
> Am 07.11.2016 um 03:10 schrieb Qu, Jim:
>> HI Christian:
>>
>>           if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS ||
>> -           amdgpu_vram_page_split == -1) {
>> +           place->lpfn || amdgpu_vram_page_split == -1) {
>>                   pages_per_node = ~0ul;
>>                   num_nodes = 1;
>>           } else {
>>
>> Here , checking place->lpfn , is that mean if required place is located in visible vram, driver run none pages split code path?
>> if so, seem it comes back to the starting point. In extreme cases, allocate visible vram fail.
>>
>> Thanks
>> JimQu
>>
>> ________________________________________
>> 发件人: amd-gfx <amd-gfx-bounces at lists.freedesktop.org> 代表 Christian König <deathsimple at vodafone.de>
>> 发送时间: 2016年11月4日 19:00
>> 收件人: amd-gfx at lists.freedesktop.org
>> 主题: [PATCH 1/2] drm/amdgpu: disable the VRAM manager on special placements v2
>>
>> From: Christian König <christian.koenig at amd.com>
>>
>> This disables the VRAM manager when a special placement is requested, otherwise
>> we play ping/pong with the buffers on every command submission.
>>
>> v2: only check lpfn
>>
>> Signed-off-by: Christian König <christian.koenig at amd.com>
>> ---
>>    drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 2 +-
>>    1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>> index 180eed7c..d710226 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c
>> @@ -108,7 +108,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
>>                   lpfn = man->size;
>>
>>           if (bo->flags & AMDGPU_GEM_CREATE_VRAM_CONTIGUOUS ||
>> -           amdgpu_vram_page_split == -1) {
>> +           place->lpfn || amdgpu_vram_page_split == -1) {
>>                   pages_per_node = ~0ul;
>>                   num_nodes = 1;
>>           } else {
>> --
>> 2.5.0
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx