[PATCH] drm/amdgpu: fix amdgpu_ttm_bo_eviction_valuable

[PATCH] drm/amdgpu: fix amdgpu_ttm_bo_eviction_valuable

[Christian König]

From: Christian König <christian.koenig@amd.com>

BOs not mapped into the GART are always valuable for an eviction. Otherwise we
don't correctly swap them out on VRAM evictions during memory pressure.

Signed-off-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
index e5d460e..ab7317e 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
@@ -1040,11 +1040,17 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
 static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 					    const struct ttm_place *place)
 {
-	if (bo->mem.mem_type == TTM_PL_VRAM &&
-	    bo->mem.start == AMDGPU_BO_INVALID_OFFSET) {
-		unsigned long num_pages = bo->mem.num_pages;
-		struct drm_mm_node *node = bo->mem.mm_node;
+	unsigned long num_pages = bo->mem.num_pages;
+	struct drm_mm_node *node = bo->mem.mm_node;
 
+	if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
+		return ttm_bo_eviction_valuable(bo, place);
+
+	switch (bo->mem.mem_type) {
+	case TTM_PL_TT:
+		return true;
+
+	case TTM_PL_VRAM:
 		/* Check each drm MM node individually */
 		while (num_pages) {
 			if (place->fpfn < (node->start + node->size) &&
@@ -1054,8 +1060,10 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
 			num_pages -= node->size;
 			++node;
 		}
+		break;
 
-		return false;
+	default:
+		break;
 	}
 
 	return ttm_bo_eviction_valuable(bo, place);
-- 
2.5.0

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH] drm/amdgpu: fix amdgpu_ttm_bo_eviction_valuable

[Alex Deucher]

On Thu, Apr 20, 2017 at 7:49 AM, Christian König
<deathsimple@vodafone.de> wrote:
> From: Christian König <christian.koenig@amd.com>
>
> BOs not mapped into the GART are always valuable for an eviction. Otherwise we
> don't correctly swap them out on VRAM evictions during memory pressure.
>
> Signed-off-by: Christian König <christian.koenig@amd.com>

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

> ---
>  drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 18 +++++++++++++-----
>  1 file changed, 13 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> index e5d460e..ab7317e 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
> @@ -1040,11 +1040,17 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
>  static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>                                             const struct ttm_place *place)
>  {
> -       if (bo->mem.mem_type == TTM_PL_VRAM &&
> -           bo->mem.start == AMDGPU_BO_INVALID_OFFSET) {
> -               unsigned long num_pages = bo->mem.num_pages;
> -               struct drm_mm_node *node = bo->mem.mm_node;
> +       unsigned long num_pages = bo->mem.num_pages;
> +       struct drm_mm_node *node = bo->mem.mm_node;
>
> +       if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
> +               return ttm_bo_eviction_valuable(bo, place);
> +
> +       switch (bo->mem.mem_type) {
> +       case TTM_PL_TT:
> +               return true;
> +
> +       case TTM_PL_VRAM:
>                 /* Check each drm MM node individually */
>                 while (num_pages) {
>                         if (place->fpfn < (node->start + node->size) &&
> @@ -1054,8 +1060,10 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>                         num_pages -= node->size;
>                         ++node;
>                 }
> +               break;
>
> -               return false;
> +       default:
> +               break;
>         }
>
>         return ttm_bo_eviction_valuable(bo, place);
> --
> 2.5.0
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH] drm/amdgpu: fix amdgpu_ttm_bo_eviction_valuable

[zhoucm1]

On 2017年04月20日 20:48, Alex Deucher wrote:
> On Thu, Apr 20, 2017 at 7:49 AM, Christian König
> <deathsimple@vodafone.de> wrote:
>> From: Christian König <christian.koenig@amd.com>
>>
>> BOs not mapped into the GART are always valuable for an eviction. Otherwise we
>> don't correctly swap them out on VRAM evictions during memory pressure.
>>
>> Signed-off-by: Christian König <christian.koenig@amd.com>
> Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Although it doesn't fix my issue, it seems more reasonable compared 
previous, Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
>
>> ---
>>   drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c | 18 +++++++++++++-----
>>   1 file changed, 13 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> index e5d460e..ab7317e 100644
>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
>> @@ -1040,11 +1040,17 @@ uint64_t amdgpu_ttm_tt_pte_flags(struct amdgpu_device *adev, struct ttm_tt *ttm,
>>   static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>>                                              const struct ttm_place *place)
>>   {
>> -       if (bo->mem.mem_type == TTM_PL_VRAM &&
>> -           bo->mem.start == AMDGPU_BO_INVALID_OFFSET) {
>> -               unsigned long num_pages = bo->mem.num_pages;
>> -               struct drm_mm_node *node = bo->mem.mm_node;
>> +       unsigned long num_pages = bo->mem.num_pages;
>> +       struct drm_mm_node *node = bo->mem.mm_node;
>>
>> +       if (bo->mem.start != AMDGPU_BO_INVALID_OFFSET)
>> +               return ttm_bo_eviction_valuable(bo, place);
>> +
>> +       switch (bo->mem.mem_type) {
>> +       case TTM_PL_TT:
>> +               return true;
>> +
>> +       case TTM_PL_VRAM:
>>                  /* Check each drm MM node individually */
>>                  while (num_pages) {
>>                          if (place->fpfn < (node->start + node->size) &&
>> @@ -1054,8 +1060,10 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct ttm_buffer_object *bo,
>>                          num_pages -= node->size;
>>                          ++node;
>>                  }
>> +               break;
>>
>> -               return false;
>> +       default:
>> +               break;
>>          }
>>
>>          return ttm_bo_eviction_valuable(bo, place);
>> --
>> 2.5.0
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx@lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx