Skip to contentSkip to conversation

drm/amdgpu: fix root reservation in amdgpu_vm_handle_fault

PATCH Superseded Pierre-Eric Pelloux-Prayer 3 participants 4 messages Apr 14

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdgpu: fix root reservation in amdgpu_vm_handle_fault

svm_range_restore_pages might reserve the root bo so it must be called after unreserving it.

The code checking that the VM still exists can be moved in the "if" block, since the VM can only be removed when the root bo is not reserved.

Fixes 32b486e8541c ("drm/amdgpu: extract amdgpu_vm_lock_by_pasid from amdgpu_vm_handle_fault")
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +15 -20
index 63156289ae7f..d86be0108913 100644
@@ -2975,25 +2975,12 @@ struct amdgpu_vm *amdgpu_vm_lock_by_pasid(struct amdgpu_device *adev,
 		return NULL;
 
 	r = amdgpu_bo_reserve(*root, true);
-	if (r)
-		goto error_unref;
-
-	/* Double check that the VM still exists */
-	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
-	vm = xa_load(&adev->vm_manager.pasids, pasid);
-	if (vm && vm->root.bo != *root)
-		vm = NULL;
-	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
-	if (!vm)
-		goto error_unlock;
+	if (r) {
+		amdgpu_bo_unref(root);
+		return NULL;
+	}
 
 	return vm;
-error_unlock:
-	amdgpu_bo_unreserve(*root);
-
-error_unref:
-	amdgpu_bo_unref(root);
-	return NULL;
 }
 
 /**
@@ -3026,11 +3013,19 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 
 	is_compute_context = vm->is_compute_context;
 
-	if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
-	    node_id, addr >> PAGE_SHIFT, ts, write_fault)) {
+	if (is_compute_context) {
+		/* Unreserve root since svm_range_restore_pages might try to reserve it. */
 		amdgpu_bo_unreserve(root);
 		amdgpu_bo_unref(&root);
-		return true;
+
+		if (!svm_range_restore_pages(adev, pasid, vmid,
+					     node_id, addr >> PAGE_SHIFT, ts, write_fault))
+			return true;
+
+		/* Double check that the VM still exists. */
+		vm = amdgpu_vm_lock_by_pasid(adev, &root, pasid);
+		if (!vm)
+			return false;
 	}
 
 	addr /= AMDGPU_GPU_PAGE_SIZE;
Signature · 1 line
2.43.0
maintainer
#1 lore raw
Quoted 6 lines · Pierre-Eric Pelloux-Prayer

svm_range_restore_pages might reserve the root bo so it must be called after unreserving it.

The code checking that the VM still exists can be moved in the "if" block, since the VM can only be removed when the root bo is not reserved.

That won't work like this. Dropping and reacquiring the root BO lock is a pretty big nono.

I think we need to fix svm_range_restore_pages() instead.

Regards,
Christian.

Quoted 64 lines
Fixes 32b486e8541c ("drm/amdgpu: extract amdgpu_vm_lock_by_pasid from amdgpu_vm_handle_fault")
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 35 +++++++++++---------------
 1 file changed, 15 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +15 -20
index 63156289ae7f..d86be0108913 100644
@@ -2975,25 +2975,12 @@ struct amdgpu_vm *amdgpu_vm_lock_by_pasid(struct amdgpu_device *adev,
 		return NULL;
 
 	r = amdgpu_bo_reserve(*root, true);
-	if (r)
-		goto error_unref;
-
-	/* Double check that the VM still exists */
-	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
-	vm = xa_load(&adev->vm_manager.pasids, pasid);
-	if (vm && vm->root.bo != *root)
-		vm = NULL;
-	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
-	if (!vm)
-		goto error_unlock;
+	if (r) {
+		amdgpu_bo_unref(root);
+		return NULL;
+	}
 
 	return vm;
-error_unlock:
-	amdgpu_bo_unreserve(*root);
-
-error_unref:
-	amdgpu_bo_unref(root);
-	return NULL;
 }
 
 /**
@@ -3026,11 +3013,19 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 
 	is_compute_context = vm->is_compute_context;
 
-	if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
-	    node_id, addr >> PAGE_SHIFT, ts, write_fault)) {
+	if (is_compute_context) {
+		/* Unreserve root since svm_range_restore_pages might try to reserve it. */
 		amdgpu_bo_unreserve(root);
 		amdgpu_bo_unref(&root);
-		return true;
+
+		if (!svm_range_restore_pages(adev, pasid, vmid,
+					     node_id, addr >> PAGE_SHIFT, ts, write_fault))
+			return true;
+
+		/* Double check that the VM still exists. */
+		vm = amdgpu_vm_lock_by_pasid(adev, &root, pasid);
+		if (!vm)
+			return false;
 	}
 
 	addr /= AMDGPU_GPU_PAGE_SIZE;
#2 lore raw

Le 14/04/2026 à 15:30, Christian König a écrit :

Quoted 9 lines · Pierre-Eric Pelloux-Prayer
Quoted 6 lines

svm_range_restore_pages might reserve the root bo so it must be called after unreserving it.

The code checking that the VM still exists can be moved in the "if" block, since the VM can only be removed when the root bo is not reserved.

That won't work like this. Dropping and reacquiring the root BO lock is a pretty big nono.

OK.

Quoted 2 lines

I think we need to fix svm_range_restore_pages() instead.

Alternatively I can modify the code from amdgpu_vm_lock_by_pasid to only reserve root when "vm->is_compute_context" is false (or to never reserve root and leave that to the caller).

Pierre-Eric

Quoted 68 lines

Regards,
Christian.

Quoted 64 lines
Fixes 32b486e8541c ("drm/amdgpu: extract amdgpu_vm_lock_by_pasid from amdgpu_vm_handle_fault")
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 35 +++++++++++---------------
  1 file changed, 15 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +15 -20
index 63156289ae7f..d86be0108913 100644
@@ -2975,25 +2975,12 @@ struct amdgpu_vm *amdgpu_vm_lock_by_pasid(struct amdgpu_device *adev,
  		return NULL;
 
  	r = amdgpu_bo_reserve(*root, true);
-	if (r)
-		goto error_unref;
-
-	/* Double check that the VM still exists */
-	xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
-	vm = xa_load(&adev->vm_manager.pasids, pasid);
-	if (vm && vm->root.bo != *root)
-		vm = NULL;
-	xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
-	if (!vm)
-		goto error_unlock;
+	if (r) {
+		amdgpu_bo_unref(root);
+		return NULL;
+	}
 
  	return vm;
-error_unlock:
-	amdgpu_bo_unreserve(*root);
-
-error_unref:
-	amdgpu_bo_unref(root);
-	return NULL;
  }
 
  /**
@@ -3026,11 +3013,19 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
 
  	is_compute_context = vm->is_compute_context;
 
-	if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
-	    node_id, addr >> PAGE_SHIFT, ts, write_fault)) {
+	if (is_compute_context) {
+		/* Unreserve root since svm_range_restore_pages might try to reserve it. */
  		amdgpu_bo_unreserve(root);
  		amdgpu_bo_unref(&root);
-		return true;
+
+		if (!svm_range_restore_pages(adev, pasid, vmid,
+					     node_id, addr >> PAGE_SHIFT, ts, write_fault))
+			return true;
+
+		/* Double check that the VM still exists. */
+		vm = amdgpu_vm_lock_by_pasid(adev, &root, pasid);
+		if (!vm)
+			return false;
  	}
 
  	addr /= AMDGPU_GPU_PAGE_SIZE;
maintainer
#3 lore raw
Quoted 11 lines · Pierre-Eric Pelloux-Prayer

Le 14/04/2026 à 15:30, Christian König a écrit :

Quoted 9 lines · Pierre-Eric Pelloux-Prayer
Quoted 6 lines

svm_range_restore_pages might reserve the root bo so it must be called after unreserving it.

The code checking that the VM still exists can be moved in the "if" block, since the VM can only be removed when the root bo is not reserved.

That won't work like this. Dropping and reacquiring the root BO lock is a pretty big nono.

Quoted 7 lines

OK.

Quoted 2 lines

I think we need to fix svm_range_restore_pages() instead.

Alternatively I can modify the code from amdgpu_vm_lock_by_pasid to only reserve root when "vm->is_compute_context" is false (or to never reserve root and leave that to the caller).

That sounds pretty awful as well.

General sequence *must* be:

xa_lock_irqsave(&pasids...);
vm = xa_load(...);
root = amdgpu_bo_ref(vm->root.bo);
xa_unlock_irqrestore(&pasid);

amdgpu_bo_reserve(root);
/* Double check that VM is still valid while holding root lock */ vm = xa_load(...);

if (!vm) {
	amdgpu_bo_unreservet(root);
	amdgpu_bo_unref(root);
	return;
}

go on update the page tables...
unlock(root);
unref(root);

When the svm handling doesn't expects the root BO to be locked (because for example it needs to lock other BOs) then we need to unlock the root BO before calling the svm functions.

Regards,
Christian.

Quoted 65 lines

Pierre-Eric

Quoted 62 lines

Regards,
Christian.

Quoted 58 lines
Fixes 32b486e8541c ("drm/amdgpu: extract amdgpu_vm_lock_by_pasid from amdgpu_vm_handle_fault")

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 35 +++++++++++---------------

  1 file changed, 15 insertions(+), 20 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +15 -19
index 63156289ae7f..d86be0108913 100644
@@ -2975,25 +2975,12 @@ struct amdgpu_vm *amdgpu_vm_lock_by_pasid(struct amdgpu_device *adev,
           return NULL;
         r = amdgpu_bo_reserve(*root, true);
-    if (r)
-        goto error_unref;
-
-    /* Double check that the VM still exists */
-    xa_lock_irqsave(&adev->vm_manager.pasids, irqflags);
-    vm = xa_load(&adev->vm_manager.pasids, pasid);
-    if (vm && vm->root.bo != *root)
-        vm = NULL;
-    xa_unlock_irqrestore(&adev->vm_manager.pasids, irqflags);
-    if (!vm)
-        goto error_unlock;
+    if (r) {
+        amdgpu_bo_unref(root);
+        return NULL;
+    }
         return vm;
-error_unlock:
-    amdgpu_bo_unreserve(*root);
-
-error_unref:
-    amdgpu_bo_unref(root);
-    return NULL;
   }
     /**
@@ -3026,11 +3013,19 @@ bool amdgpu_vm_handle_fault(struct amdgpu_device *adev, u32 pasid,
         is_compute_context = vm->is_compute_context;
   -    if (is_compute_context && !svm_range_restore_pages(adev, pasid, vmid,
-        node_id, addr >> PAGE_SHIFT, ts, write_fault)) {
+    if (is_compute_context) {
+        /* Unreserve root since svm_range_restore_pages might try to reserve it. */
           amdgpu_bo_unreserve(root);
           amdgpu_bo_unref(&root);
-        return true;
+
+        if (!svm_range_restore_pages(adev, pasid, vmid,
+                         node_id, addr >> PAGE_SHIFT, ts, write_fault))
+            return true;
+
+        /* Double check that the VM still exists. */
+        vm = amdgpu_vm_lock_by_pasid(adev, &root, pasid);
+        if (!vm)
+            return false;
       }
         addr /= AMDGPU_GPU_PAGE_SIZE;