Skip to contentSkip to conversation

drm/amd/amdgpu: Bail out of BO node creation if not enough VRAM

PATCH In review StDenis, Tom 2 participants 3 messages Jun 10, 2019 – Jun 11, 2019 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amd/amdgpu: Bail out of BO node creation if not enough VRAM

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +12 -5
index 8aea2f21b202..70b4a5a97ed2 100644
@@ -276,7 +276,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 	struct drm_mm_node *nodes;
 	enum drm_mm_insert_mode mode;
 	unsigned long lpfn, num_nodes, pages_per_node, pages_left;
-	uint64_t usage = 0, vis_usage = 0;
+	uint64_t vis_usage = 0;
 	unsigned i;
 	int r;
 
@@ -284,6 +284,13 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 	if (!lpfn)
 		lpfn = man->size;
 
+	/* bail out quickly if there's likely not enough VRAM for this BO */
+	atomic64_add(mem->num_pages << PAGE_SHIFT, &mgr->usage);
+	if (atomic64_read(&mgr->usage) > adev->gmc.mc_vram_size) {
+		atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
+		return -ENOSPC;
+	}
+
 	if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
 		pages_per_node = ~0ul;
 		num_nodes = 1;
@@ -300,8 +307,10 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 
 	nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
 			       GFP_KERNEL | __GFP_ZERO);
-	if (!nodes)
+	if (!nodes) {
+		atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
 		return -ENOMEM;
+	}
 
 	mode = DRM_MM_INSERT_BEST;
 	if (place->flags & TTM_PL_FLAG_TOPDOWN)
@@ -321,7 +330,6 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 		if (unlikely(r))
 			break;
 
-		usage += nodes[i].size << PAGE_SHIFT;
 		vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
 		amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
 		pages_left -= pages;
@@ -341,14 +349,12 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 		if (unlikely(r))
 			goto error;
 
-		usage += nodes[i].size << PAGE_SHIFT;
 		vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
 		amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
 		pages_left -= pages;
 	}
 	spin_unlock(&mgr->lock);
 
-	atomic64_add(usage, &mgr->usage);
 	atomic64_add(vis_usage, &mgr->vis_usage);
 
 	mem->mm_node = nodes;
@@ -359,6 +365,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 	while (i--)
 		drm_mm_remove_node(&nodes[i]);
 	spin_unlock(&mgr->lock);
+	atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
 
 	kvfree(nodes);
 	return r == -ENOSPC ? 0 : r;
Signature · 6 lines
2.21.0

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): [email protected]

#2 lore raw

Am 10.06.19 um 16:32 schrieb StDenis, Tom:

Quoted 19 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 17 ++++++++++++-----
  1 file changed, 12 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +1 -1
index 8aea2f21b202..70b4a5a97ed2 100644
@@ -276,7 +276,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  	struct drm_mm_node *nodes;
  	enum drm_mm_insert_mode mode;
  	unsigned long lpfn, num_nodes, pages_per_node, pages_left;
-	uint64_t usage = 0, vis_usage = 0;
+	uint64_t vis_usage = 0;
  	unsigned i;
  	int r;
 

drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, · line 292

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +6 -0
@@ -284,6 +284,13 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  	if (!lpfn)
  		lpfn = man->size;
 
+	/* bail out quickly if there's likely not enough VRAM for this BO */
+	atomic64_add(mem->num_pages << PAGE_SHIFT, &mgr->usage);
+	if (atomic64_read(&mgr->usage) > adev->gmc.mc_vram_size) {
+		atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
+		return -ENOSPC;
+	}

That's not 100% correct IIRC. See TTM expects a return value of 0, but mem->node to be NULL when allocation fails.

Christian.

Quoted 40 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +4 -4
+
  	if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
  		pages_per_node = ~0ul;
  		num_nodes = 1;
@@ -300,8 +307,10 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
 
  	nodes = kvmalloc_array((uint32_t)num_nodes, sizeof(*nodes),
  			       GFP_KERNEL | __GFP_ZERO);
-	if (!nodes)
+	if (!nodes) {
+		atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
  		return -ENOMEM;
+	}
 
  	mode = DRM_MM_INSERT_BEST;
  	if (place->flags & TTM_PL_FLAG_TOPDOWN)
@@ -321,7 +330,6 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  		if (unlikely(r))
  			break;
 
-		usage += nodes[i].size << PAGE_SHIFT;
  		vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
  		amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
  		pages_left -= pages;
@@ -341,14 +349,12 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  		if (unlikely(r))
  			goto error;
 
-		usage += nodes[i].size << PAGE_SHIFT;
  		vis_usage += amdgpu_vram_mgr_vis_size(adev, &nodes[i]);
  		amdgpu_vram_mgr_virt_start(mem, &nodes[i]);
  		pages_left -= pages;
  	}
  	spin_unlock(&mgr->lock);
 
-	atomic64_add(usage, &mgr->usage);
  	atomic64_add(vis_usage, &mgr->vis_usage);
 
  	mem->mm_node = nodes;

drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, · line 371

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +1 -0
@@ -359,6 +365,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  	while (i--)
  		drm_mm_remove_node(&nodes[i]);
  	spin_unlock(&mgr->lock);
+	atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
 
  	kvfree(nodes);
  	return r == -ENOSPC ? 0 : r;

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 91bb2355-949f-43a3-67e6-fb752de86b5c-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#4 lore raw
Quoted 26 lines · Christian König

Am 10.06.19 um 16:32 schrieb StDenis, Tom:

Quoted 25 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 17 ++++++++++++-----

  1 file changed, 12 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +1 -1
index 8aea2f21b202..70b4a5a97ed2 100644
@@ -276,7 +276,7 @@ static int amdgpu_vram_mgr_new(struct 
 ttm_mem_type_manager *man,
       struct drm_mm_node *nodes;
       enum drm_mm_insert_mode mode;
       unsigned long lpfn, num_nodes, pages_per_node, pages_left;
-    uint64_t usage = 0, vis_usage = 0;
+    uint64_t vis_usage = 0;
       unsigned i;
       int r;

  @@ -284,6 +284,13 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
      if (!lpfn)
          lpfn = man->size;
  +    /* bail out quickly if there's likely not enough VRAM for this BO */

Quoted 8 lines
Quoted 5 lines

+    atomic64_add(mem->num_pages << PAGE_SHIFT, &mgr->usage); +    if (atomic64_read(&mgr->usage) > adev->gmc.mc_vram_size) { +        atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage); +        return -ENOSPC;
+    }

That's not 100% correct IIRC. See TTM expects a return value of 0, but mem->node to be NULL when allocation fails.

Yup, the node creation returns -ENOSPC which this function then translates to 0.  I'll submit a v2 in a minute once I pass a piglit test :-)

Tom

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx