Skip to contentSkip to conversation

drm/amdgpu: fix nullptr error of amdgpu_vm_handle_moved

PATCH In review Heng Zhou 4 participants 5 messages Aug 8, 2025 – Aug 11, 2025

Subsystems (3)
lore
Share & tools
lore

Heng Zhou

author
#0 lore raw

PATCH 1/1 drm/amdgpu: fix nullptr error of amdgpu_vm_handle_moved

If a amdgpu_bo_va is fpriv->prt_va, the bo of this one is always NULL. So, such kind of amdgpu_bo_va should be updated separately before amdgpu_vm_handle_moved.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           | 8 ++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h           | 1 +
 3 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -0
index 37d8a7034a7e..e795b2970620 100644
@@ -2970,6 +2970,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 		struct amdgpu_device *adev = amdgpu_ttm_adev(
 			peer_vm->root.bo->tbo.bdev);
 
+		ret = amdgpu_vm_handle_prt_moved(adev,peer_vm);
+		if (ret) {
+			pr_debug("Memory eviction: handle PRT moved failed. Try again\n");
+			goto validate_map_fail;
+		}
+
 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
 		if (ret) {
 			pr_debug("Memory eviction: handle moved failed. Try again\n");
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -0
index 283dd44f04b0..2c2a93f22ba0 100644
@@ -1569,6 +1569,14 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 
 }
 
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	struct amdgpu_fpriv *fpriv;
+
+	fpriv = container_of(vm, struct amdgpu_fpriv, vm);
+	return amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
+}
+
 /**
  * amdgpu_vm_handle_moved - handle moved BOs in the PT
  *
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index f9549f6b3d1f..853a66f9fd2e 100644
@@ -504,6 +504,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);
Signature · 1 line
2.43.0
#1 lore raw
Quoted 16 lines · Heng Zhou

If a amdgpu_bo_va is fpriv->prt_va, the bo of this one is always NULL. So, such kind of amdgpu_bo_va should be updated separately before amdgpu_vm_handle_moved.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           | 8 ++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h           | 1 +
 3 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +0 -0
index 37d8a7034a7e..e795b2970620 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu * · line 2975

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -0
@@ -2970,6 +2970,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 		struct amdgpu_device *adev = amdgpu_ttm_adev(
 			peer_vm->root.bo->tbo.bdev);
 
+		ret = amdgpu_vm_handle_prt_moved(adev,peer_vm);
+		if (ret) {
+			pr_debug("Memory eviction: handle PRT moved failed. Try again\n");

Though this follows the existing format, better to change this dev_dbg and add more details like task_info/pid from the vm.

Thanks,
Lijo

Quoted 37 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +3 -0
+			goto validate_map_fail;
+		}
+
 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
 		if (ret) {
 			pr_debug("Memory eviction: handle moved failed. Try again\n");
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +8 -0
index 283dd44f04b0..2c2a93f22ba0 100644
@@ -1569,6 +1569,14 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 
 }
 
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	struct amdgpu_fpriv *fpriv;
+
+	fpriv = container_of(vm, struct amdgpu_fpriv, vm);
+	return amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
+}
+
 /**
  * amdgpu_vm_handle_moved - handle moved BOs in the PT
  *
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index f9549f6b3d1f..853a66f9fd2e 100644
@@ -504,6 +504,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);
maintainer
#2 lore raw
In reply to Heng Zhou
Quoted 33 lines · Heng Zhou

If a amdgpu_bo_va is fpriv->prt_va, the bo of this one is always NULL. So, such kind of amdgpu_bo_va should be updated separately before amdgpu_vm_handle_moved.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           | 8 ++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h           | 1 +
 3 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -0
index 37d8a7034a7e..e795b2970620 100644
@@ -2970,6 +2970,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 		struct amdgpu_device *adev = amdgpu_ttm_adev(
 			peer_vm->root.bo->tbo.bdev);
 
+		ret = amdgpu_vm_handle_prt_moved(adev,peer_vm);
+		if (ret) {
+			pr_debug("Memory eviction: handle PRT moved failed. Try again\n");
+			goto validate_map_fail;
+		}
+
 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
 		if (ret) {
 			pr_debug("Memory eviction: handle moved failed. Try again\n");
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 283dd44f04b0..2c2a93f22ba0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_clear_freed(struct amdgpu_device *adev, · line 1576

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +5 -0
@@ -1569,6 +1569,14 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 
 }
 
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	struct amdgpu_fpriv *fpriv;
+
+	fpriv = container_of(vm, struct amdgpu_fpriv, vm);

That's an absolute no-go from inside the VM code.

You could maybe do that in the KFD code, but Felix and Philip need to decide that.

Regards,
Christian.

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
+	return amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
+}
+
 /**
  * amdgpu_vm_handle_moved - handle moved BOs in the PT
  *
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index f9549f6b3d1f..853a66f9fd2e 100644
@@ -504,6 +504,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);
maintainer
#3 lore raw
Quoted 38 lines · Christian König
Quoted 36 lines · Heng Zhou

If a amdgpu_bo_va is fpriv->prt_va, the bo of this one is always NULL. So, such kind of amdgpu_bo_va should be updated separately before amdgpu_vm_handle_moved.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           | 8 ++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h           | 1 +
 3 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -0
index 37d8a7034a7e..e795b2970620 100644
@@ -2970,6 +2970,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 		struct amdgpu_device *adev = amdgpu_ttm_adev(
 			peer_vm->root.bo->tbo.bdev);
 
+		ret = amdgpu_vm_handle_prt_moved(adev,peer_vm);
+		if (ret) {
+			pr_debug("Memory eviction: handle PRT moved failed. Try again\n");
+			goto validate_map_fail;
+		}
+
 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
 		if (ret) {
 			pr_debug("Memory eviction: handle moved failed. Try again\n");
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 283dd44f04b0..2c2a93f22ba0 100644
@@ -1569,6 +1569,14 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 
 }
 
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_clear_freed(struct amdgpu_device *adev, · line 1576

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +5 -0
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	struct amdgpu_fpriv *fpriv;
+
+	fpriv = container_of(vm, struct amdgpu_fpriv, vm);

That's an absolute no-go from inside the VM code.

You could maybe do that in the KFD code, but Felix and Philip need to decide that.

We're calling amdgpu_vm_handle_moved in amdgpu_amdkfd_gpuvm_restore_process_bos to update mappings of BOs that are not managed by KFD. I don't think we care about PRT BOs, though. They are not useful for compute work.

What is the problem you're trying to fix here? Is amdgpu_vm_handle_moved crashing because of the PRT BO? In that case I'd try to make amdgpu_vm_handle_moved more robust to handle that situation gracefully (i.e. ignoring PRT BOs).

Regards,
  Felix

Quoted 22 lines

Regards,
Christian.

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
+	return amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
+}
+
 /**
  * amdgpu_vm_handle_moved - handle moved BOs in the PT
  *
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index f9549f6b3d1f..853a66f9fd2e 100644
@@ -504,6 +504,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);
maintainer
#4 lore raw
Quoted 47 lines · Felix Kuehling
Quoted 45 lines · Christian König
Quoted 33 lines · Heng Zhou

If a amdgpu_bo_va is fpriv->prt_va, the bo of this one is always NULL. So, such kind of amdgpu_bo_va should be updated separately before amdgpu_vm_handle_moved.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 6 ++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c           | 8 ++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h           | 1 +
 3 files changed, 15 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +6 -0
index 37d8a7034a7e..e795b2970620 100644
@@ -2970,6 +2970,12 @@ int amdgpu_amdkfd_gpuvm_restore_process_bos(void *info, struct dma_fence __rcu *
 		struct amdgpu_device *adev = amdgpu_ttm_adev(
 			peer_vm->root.bo->tbo.bdev);
 
+		ret = amdgpu_vm_handle_prt_moved(adev,peer_vm);
+		if (ret) {
+			pr_debug("Memory eviction: handle PRT moved failed. Try again\n");
+			goto validate_map_fail;
+		}
+
 		ret = amdgpu_vm_handle_moved(adev, peer_vm, &exec.ticket);
 		if (ret) {
 			pr_debug("Memory eviction: handle moved failed. Try again\n");
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index 283dd44f04b0..2c2a93f22ba0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_clear_freed(struct amdgpu_device *adev, · line 1576

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +5 -0
@@ -1569,6 +1569,14 @@ int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 
 }
 
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm)
+{
+	struct amdgpu_fpriv *fpriv;
+
+	fpriv = container_of(vm, struct amdgpu_fpriv, vm);

That's an absolute no-go from inside the VM code.

Quoted 5 lines
Quoted 1 line

You could maybe do that in the KFD code, but Felix and Philip need to decide that.

We're calling amdgpu_vm_handle_moved in amdgpu_amdkfd_gpuvm_restore_process_bos to update mappings of BOs that are not managed by KFD. I don't think we care about PRT BOs, though. They are not useful for compute work.

What is the problem you're trying to fix here? Is amdgpu_vm_handle_moved crashing because of the PRT BO? In that case I'd try to make amdgpu_vm_handle_moved more robust to handle that situation gracefully (i.e. ignoring PRT BOs).

The issue is that amdgpu_vm_handle_moved() only takes care of moved BOs, but not PRT mappings since they can never "move" because they don't have an associated BO.

But after a GPU reset the all mappings are marked invalid to re-create the page tables, so the PRT mappings ended up on the "moved" list as well and that is not handled by amdgpu_vm_handle_moved() so you end up in a crash.

But the real question is how we end up in this situation in the first place? A GPU reset should kill the KFD process, so there should never be a restore attempted.

Regards,
Christian.

Quoted 27 lines

Regards,
  Felix

Quoted 22 lines

Regards,
Christian.

Quoted 18 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
+	return amdgpu_vm_bo_update(adev, fpriv->prt_va, false);
+}
+
 /**
  * amdgpu_vm_handle_moved - handle moved BOs in the PT
  *
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +1 -0
index f9549f6b3d1f..853a66f9fd2e 100644
@@ -504,6 +504,7 @@ int amdgpu_vm_update_pdes(struct amdgpu_device *adev,
 int amdgpu_vm_clear_freed(struct amdgpu_device *adev,
 			  struct amdgpu_vm *vm,
 			  struct dma_fence **fence);
+int amdgpu_vm_handle_prt_moved(struct amdgpu_device *adev, struct amdgpu_vm *vm);
 int amdgpu_vm_handle_moved(struct amdgpu_device *adev,
 			   struct amdgpu_vm *vm,
 			   struct ww_acquire_ctx *ticket);