Skip to contentSkip to conversation

drm/amd/amdgpu: Bail out of BO node creation if not enough VRAM (v2)

PATCH In review StDenis, Tom 3 participants 4 messages Jun 11, 2019 – Jun 12, 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 (v2)

(v2): Return 0 and set mem->mm_node to NULL.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +13 -5
index 8aea2f21b202..c2730e5c1081 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 +284,14 @@ 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);
+		mem->mm_node = NULL;
+		return 0;
+	}
+
 	if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
 		pages_per_node = ~0ul;
 		num_nodes = 1;
@@ -300,8 +308,10 @@ 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 +331,6 @@ 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 +350,12 @@ 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 +366,7 @@ 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
Quoted 10 lines · Tom

(v2): Return 0 and set mem->mm_node to NULL.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 18 +++++++++++++-----
 1 file changed, 13 insertions(+), 5 deletions(-)

[...]

drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, · line 289

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +3 -0
@@ -284,6 +284,14 @@ 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) {

Should probably use atomic64_add_return instead of atomic64_add + atomic64_read.

Also, AFAICT this doesn't allow any VRAM overcommit, which seems a bit conservative?

Signature · 6 lines
Earthling Michel Dänzer               |              https://www.amd.com
Libre software enthusiast             |             Mesa and X developer
_______________________________________________
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): fd4c5152-2670-8ac5-e6cc-fff6f260f3ed-otUistvHUpPR7s880joybQ@public.gmane.org

#4 lore raw

Am 12.06.19 um 09:26 schrieb Michel Dänzer:

Quoted 14 lines · Tom
Quoted 13 lines

(v2): Return 0 and set mem->mm_node to NULL.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 18 +++++++++++++-----
  1 file changed, 13 insertions(+), 5 deletions(-)

[...]

drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +0 -0
@@ -284,6 +284,14 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man,
  	if (!lpfn)
  		lpfn = man->size;
 
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, · line 289

Quoted 3 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +3 -0
+	/* 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) {

Should probably use atomic64_add_return instead of atomic64_add + atomic64_read.

Also, AFAICT this doesn't allow any VRAM overcommit, which seems a bit conservative?

This is the low level VRAM manager. Here we can't overcomit anyway and need to return early for TTM to evict things.

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

#5 lore raw
In reply to a message not in the archive

Am 11.06.19 um 18:54 schrieb StDenis, Tom:

Quoted 21 lines

(v2): Return 0 and set mem->mm_node to NULL.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 18 +++++++++++++-----
  1 file changed, 13 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +1 -1
index 8aea2f21b202..c2730e5c1081 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 289

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +3 -0
@@ -284,6 +284,14 @@ 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) {

You should use atomic64_add_return here instead of two operations. Otherwise using an atomic doesn't make to much sense.

Christian.

Quoted 44 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +8 -4
+		atomic64_sub(mem->num_pages << PAGE_SHIFT, &mgr->usage);
+		mem->mm_node = NULL;
+		return 0;
+	}
+
  	if (place->flags & TTM_PL_FLAG_CONTIGUOUS) {
  		pages_per_node = ~0ul;
  		num_nodes = 1;
@@ -300,8 +308,10 @@ 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 +331,6 @@ 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 +350,12 @@ 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 372

Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c +1 -0
@@ -359,6 +366,7 @@ 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