drm/amdgpu: stop using addr to check for BO move v3

It is theoretically possible that a swapped out BO gets the
same GTT address, but different backing pages while being swapped in.

Instead just use another VA state to note updated areas.
Ported from not upstream yet radeon commit with the same name.

v2: fix some bugs in the original implementation found in the radeon code.
v3: squash in VCE/UVD fix
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu.h

@@ -539,14 +539,16 @@ struct amdgpu_bo_va_mapping {
 struct amdgpu_bo_va {
 	/* protected by bo being reserved */
 	struct list_head		bo_list;
-	uint64_t			addr;
 	struct amdgpu_fence		*last_pt_update;
 	unsigned			ref_count;
 
-	/* protected by vm mutex */
-	struct list_head		mappings;
+	/* protected by vm mutex and spinlock */
 	struct list_head		vm_status;
 
+	/* mappings for this bo_va */
+	struct list_head		invalids;
+	struct list_head		valids;
+
 	/* constant after initialization */
 	struct amdgpu_vm		*vm;
 	struct amdgpu_bo		*bo;
@@ -964,13 +966,16 @@ struct amdgpu_vm {
 
 	struct rb_root		va;
 
-	/* protecting invalidated and freed */
+	/* protecting invalidated */
 	spinlock_t		status_lock;
 
 	/* BOs moved, but not yet updated in the PT */
 	struct list_head	invalidated;
 
-	/* BOs freed, but not yet updated in the PT */
+	/* BOs cleared in the PT because of a move */
+	struct list_head	cleared;
+
+	/* BO mappings freed, but not yet updated in the PT */
 	struct list_head	freed;
 
 	/* contains the page directory */

drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c

@@ -1009,7 +1009,16 @@ amdgpu_cs_find_mapping(struct amdgpu_cs_parser *parser,
 		if (!reloc->bo_va)
 			continue;
 
-		list_for_each_entry(mapping, &reloc->bo_va->mappings, list) {
+		list_for_each_entry(mapping, &reloc->bo_va->valids, list) {
+			if (mapping->it.start > addr ||
+			    addr > mapping->it.last)
+				continue;
+
+			*bo = reloc->bo_va->bo;
+			return mapping;
+		}
+
+		list_for_each_entry(mapping, &reloc->bo_va->invalids, list) {
 			if (mapping->it.start > addr ||
 			    addr > mapping->it.last)
 				continue;

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

@@ -951,21 +951,24 @@ int amdgpu_vm_bo_update(struct amdgpu_device *adev,
 		addr = 0;
 	}
 
-	if (addr == bo_va->addr)
-		return 0;
-
 	flags = amdgpu_ttm_tt_pte_flags(adev, bo_va->bo->tbo.ttm, mem);
 
-	list_for_each_entry(mapping, &bo_va->mappings, list) {
+	spin_lock(&vm->status_lock);
+	if (!list_empty(&bo_va->vm_status))
+		list_splice_init(&bo_va->valids, &bo_va->invalids);
+	spin_unlock(&vm->status_lock);
+
+	list_for_each_entry(mapping, &bo_va->invalids, list) {
 		r = amdgpu_vm_bo_update_mapping(adev, vm, mapping, addr,
 						flags, &bo_va->last_pt_update);
 		if (r)
 			return r;
 	}
 
-	bo_va->addr = addr;
 	spin_lock(&vm->status_lock);
 	list_del_init(&bo_va->vm_status);
+	if (!mem)
+		list_add(&bo_va->vm_status, &vm->cleared);
 	spin_unlock(&vm->status_lock);
 
 	return 0;
@@ -1065,10 +1068,10 @@ struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev,
 	}
 	bo_va->vm = vm;
 	bo_va->bo = bo;
-	bo_va->addr = 0;
 	bo_va->ref_count = 1;
 	INIT_LIST_HEAD(&bo_va->bo_list);
-	INIT_LIST_HEAD(&bo_va->mappings);
+	INIT_LIST_HEAD(&bo_va->valids);
+	INIT_LIST_HEAD(&bo_va->invalids);
 	INIT_LIST_HEAD(&bo_va->vm_status);
 
 	mutex_lock(&vm->mutex);
@@ -1157,12 +1160,10 @@ int amdgpu_vm_bo_map(struct amdgpu_device *adev,
 	mapping->offset = offset;
 	mapping->flags = flags;
 
-	list_add(&mapping->list, &bo_va->mappings);
+	list_add(&mapping->list, &bo_va->invalids);
 	interval_tree_insert(&mapping->it, &vm->va);
 	trace_amdgpu_vm_bo_map(bo_va, mapping);
 
-	bo_va->addr = 0;
-
 	/* Make sure the page tables are allocated */
 	saddr >>= amdgpu_vm_block_size;
 	eaddr >>= amdgpu_vm_block_size;
@@ -1243,30 +1244,38 @@ int amdgpu_vm_bo_unmap(struct amdgpu_device *adev,
 {
 	struct amdgpu_bo_va_mapping *mapping;
 	struct amdgpu_vm *vm = bo_va->vm;
+	bool valid = true;
 
 	saddr /= AMDGPU_GPU_PAGE_SIZE;
 
-	list_for_each_entry(mapping, &bo_va->mappings, list) {
+	list_for_each_entry(mapping, &bo_va->valids, list) {
+		if (mapping->it.start == saddr)
+			break;
+	}
+
+	if (&mapping->list == &bo_va->valids) {
+		valid = false;
+
+		list_for_each_entry(mapping, &bo_va->invalids, list) {
 			if (mapping->it.start == saddr)
 				break;
 		}
 
-	if (&mapping->list == &bo_va->mappings) {
+		if (&mapping->list == &bo_va->invalids) {
 			amdgpu_bo_unreserve(bo_va->bo);
 			return -ENOENT;
 		}
+	}
 
 	mutex_lock(&vm->mutex);
 	list_del(&mapping->list);
 	interval_tree_remove(&mapping->it, &vm->va);
 	trace_amdgpu_vm_bo_unmap(bo_va, mapping);
 
-	if (bo_va->addr) {
-		/* clear the old address */
+	if (valid)
 		list_add(&mapping->list, &vm->freed);
-	} else {
+	else
 		kfree(mapping);
-	}
 	mutex_unlock(&vm->mutex);
 	amdgpu_bo_unreserve(bo_va->bo);
 
@@ -1297,15 +1306,18 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev,
 	list_del(&bo_va->vm_status);
 	spin_unlock(&vm->status_lock);
 
-	list_for_each_entry_safe(mapping, next, &bo_va->mappings, list) {
+	list_for_each_entry_safe(mapping, next, &bo_va->valids, list) {
 		list_del(&mapping->list);
 		interval_tree_remove(&mapping->it, &vm->va);
 		trace_amdgpu_vm_bo_unmap(bo_va, mapping);
-		if (bo_va->addr)
 		list_add(&mapping->list, &vm->freed);
-		else
+	}
+	list_for_each_entry_safe(mapping, next, &bo_va->invalids, list) {
+		list_del(&mapping->list);
+		interval_tree_remove(&mapping->it, &vm->va);
 		kfree(mapping);
 	}
+
 	amdgpu_fence_unref(&bo_va->last_pt_update);
 	kfree(bo_va);
 
@@ -1327,13 +1339,11 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev,
 	struct amdgpu_bo_va *bo_va;
 
 	list_for_each_entry(bo_va, &bo->va, bo_list) {
-		if (bo_va->addr) {
 		spin_lock(&bo_va->vm->status_lock);
-			list_del(&bo_va->vm_status);
+		if (list_empty(&bo_va->vm_status))
 			list_add(&bo_va->vm_status, &bo_va->vm->invalidated);
 		spin_unlock(&bo_va->vm->status_lock);
 	}
-	}
 }
 
 /**
@@ -1360,6 +1370,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 	vm->va = RB_ROOT;
 	spin_lock_init(&vm->status_lock);
 	INIT_LIST_HEAD(&vm->invalidated);
+	INIT_LIST_HEAD(&vm->cleared);
 	INIT_LIST_HEAD(&vm->freed);
 
 	pd_size = amdgpu_vm_directory_size(adev);