[PATCH] drm/amdgpu: try allowed domain when pin framebuffer failed

[Christian König]

Am 07.12.22 um 17:08 schrieb Alex Deucher:
> On Wed, Dec 7, 2022 at 10:52 AM Christian König
> <ckoenig.leichtzumerken at gmail.com> wrote:
>> Am 07.12.22 um 16:38 schrieb Alex Deucher:
>>> On Wed, Dec 7, 2022 at 10:23 AM Christian König
>>> <ckoenig.leichtzumerken at gmail.com> wrote:
>>>> I would go a step further and just allow GTT domain on ASICs != CARRIZO
>>>> | STONEY.
>>>>
>>>> I can't see a good reason we should still have any limitation here, VRAM
>>>> doesn't have any advantage any more as far as I know.
>>> Well, if VRAM is available we want to make sure someone uses it
>>> otherwise it's just wasted.
>> Well it still gets used when it's free. So now problem at all here.
>>
>> We should just not force anything into VRAM or GTT any more if it's
>> technically not necessary.
> So just this?
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> index 919bbea2e3ac..8e8f07fa7a93 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
> @@ -1506,7 +1506,7 @@ u64 amdgpu_bo_gpu_offset_no_check(struct amdgpu_bo *bo)
>   uint32_t amdgpu_bo_get_preferred_domain(struct amdgpu_device *adev,
>                                              uint32_t domain)
>   {
> -       if (domain == (AMDGPU_GEM_DOMAIN_VRAM | AMDGPU_GEM_DOMAIN_GTT)) {

We still need to keep this check to avoid trying to adjust VRAM only 
allocations (the cursor still needs this IIRC).

Apart from that I think that should work.

Christian.

> +       if ((adev->asic_type == CHIP_CARRIZO) || (adev->asic_type ==
> CHIP_STONEY)) {
>                  domain = AMDGPU_GEM_DOMAIN_VRAM;
>                  if (adev->gmc.real_vram_size <= AMDGPU_SG_THRESHOLD)
>                          domain = AMDGPU_GEM_DOMAIN_GTT;
>
>
>
>> Christian.
>>
>>> Alex
>>>
>>>
>>>> Christian.
>>>>
>>>> Am 07.12.22 um 16:10 schrieb Alex Deucher:
>>>>> Does this patch fix the problem?
>>>>>
>>>>> Alex
>>>>>
>>>>> On Wed, Dec 7, 2022 at 2:27 AM Zhang, Jesse(Jie) <Jesse.Zhang at amd.com> wrote:
>>>>>> [AMD Official Use Only - General]
>>>>>>
>>>>>>
>>>>>>        drm/amdgpu: try allowed domain when pin framebuffer failed.
>>>>>>
>>>>>>
>>>>>>
>>>>>>        [WHY&HOW]
>>>>>>
>>>>>>
>>>>>>
>>>>>>        in some scenarios, the allocate memory often failed. such as do hot plug or play games.
>>>>>>
>>>>>>        so we can try allowed domain, if the preferred domain cannot allocate memory.
>>>>>>
>>>>>>
>>>>>>
>>>>>>        Signed-off-by: jie1zhan jesse.zhang at amd.com
>>>>>>
>>>>>>        Change-Id: I4b62e2ff072d02c515f901000a5789339d481273
>>>>>>
>>>>>>
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>
>>>>>> index 1ae0c8723348..05fcaf7f9d92 100644
>>>>>>
>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>
>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c
>>>>>>
>>>>>> @@ -39,6 +39,7 @@
>>>>>>
>>>>>> #include "amdgpu.h"
>>>>>>
>>>>>> #include "amdgpu_trace.h"
>>>>>>
>>>>>> #include "amdgpu_amdkfd.h"
>>>>>>
>>>>>> +#include "amdgpu_display.h"
>>>>>>
>>>>>>
>>>>>>
>>>>>> /**
>>>>>>
>>>>>>      * DOC: amdgpu_object
>>>>>>
>>>>>> @@ -942,8 +943,14 @@ int amdgpu_bo_pin_restricted(struct amdgpu_bo *bo, u32 domain,
>>>>>>
>>>>>>                            bo->placements[i].lpfn = lpfn;
>>>>>>
>>>>>>            }
>>>>>>
>>>>>>
>>>>>>
>>>>>> +       retry:
>>>>>>
>>>>>>            r = ttm_bo_validate(&bo->tbo, &bo->placement, &ctx);
>>>>>>
>>>>>>            if (unlikely(r)) {
>>>>>>
>>>>>> +               //try allowed domain when pin failed. just a workaround.
>>>>>>
>>>>>> +               if (unlikely(r == -ENOMEM) && domain != bo->allowed_domains) {
>>>>>>
>>>>>> +                       amdgpu_bo_placement_from_domain(bo, bo->allowed_domains);
>>>>>>
>>>>>> +                       goto retry;
>>>>>>
>>>>>> +               }
>>>>>>
>>>>>>                    dev_err(adev->dev, "%p pin failed\n", bo);
>>>>>>
>>>>>>                    goto error;
>>>>>>
>>>>>>            }