drm/amdgpu: add some VM PD/PT iterators v2
authorChristian König <christian.koenig@amd.com>
Sat, 1 Sep 2018 08:36:48 +0000 (10:36 +0200)
committerAlex Deucher <alexander.deucher@amd.com>
Thu, 13 Sep 2018 20:14:08 +0000 (15:14 -0500)
Both a leaf as well as dfs iterator to walk over all the PDs/PTs.

v2: update comments and fix for_each_amdgpu_vm_pt_dfs_safe

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

index be1659fedf9453b11f06a26456162e2383075a90..06a173e843ceceb32bcba197ecc6c61870b835c0 100644 (file)
@@ -355,6 +355,230 @@ static struct amdgpu_vm_pt *amdgpu_vm_pt_parent(struct amdgpu_vm_pt *pt)
        return list_first_entry(&parent->va, struct amdgpu_vm_pt, base.bo_list);
 }
 
+/**
+ * amdgpu_vm_pt_cursor - state for for_each_amdgpu_vm_pt
+ */
+struct amdgpu_vm_pt_cursor {
+       uint64_t pfn;
+       struct amdgpu_vm_pt *parent;
+       struct amdgpu_vm_pt *entry;
+       unsigned level;
+};
+
+/**
+ * amdgpu_vm_pt_start - start PD/PT walk
+ *
+ * @adev: amdgpu_device pointer
+ * @vm: amdgpu_vm structure
+ * @start: start address of the walk
+ * @cursor: state to initialize
+ *
+ * Initialize a amdgpu_vm_pt_cursor to start a walk.
+ */
+static void amdgpu_vm_pt_start(struct amdgpu_device *adev,
+                              struct amdgpu_vm *vm, uint64_t start,
+                              struct amdgpu_vm_pt_cursor *cursor)
+{
+       cursor->pfn = start;
+       cursor->parent = NULL;
+       cursor->entry = &vm->root;
+       cursor->level = adev->vm_manager.root_level;
+}
+
+/**
+ * amdgpu_vm_pt_descendant - go to child node
+ *
+ * @adev: amdgpu_device pointer
+ * @cursor: current state
+ *
+ * Walk to the child node of the current node.
+ * Returns:
+ * True if the walk was possible, false otherwise.
+ */
+static bool amdgpu_vm_pt_descendant(struct amdgpu_device *adev,
+                                   struct amdgpu_vm_pt_cursor *cursor)
+{
+       unsigned num_entries, shift, idx;
+
+       if (!cursor->entry->entries)
+               return false;
+
+       BUG_ON(!cursor->entry->base.bo);
+       num_entries = amdgpu_vm_num_entries(adev, cursor->level);
+       shift = amdgpu_vm_level_shift(adev, cursor->level);
+
+       ++cursor->level;
+       idx = (cursor->pfn >> shift) % num_entries;
+       cursor->parent = cursor->entry;
+       cursor->entry = &cursor->entry->entries[idx];
+       return true;
+}
+
+/**
+ * amdgpu_vm_pt_sibling - go to sibling node
+ *
+ * @adev: amdgpu_device pointer
+ * @cursor: current state
+ *
+ * Walk to the sibling node of the current node.
+ * Returns:
+ * True if the walk was possible, false otherwise.
+ */
+static bool amdgpu_vm_pt_sibling(struct amdgpu_device *adev,
+                                struct amdgpu_vm_pt_cursor *cursor)
+{
+       unsigned shift, num_entries;
+
+       /* Root doesn't have a sibling */
+       if (!cursor->parent)
+               return false;
+
+       /* Go to our parents and see if we got a sibling */
+       shift = amdgpu_vm_level_shift(adev, cursor->level - 1);
+       num_entries = amdgpu_vm_num_entries(adev, cursor->level - 1);
+
+       if (cursor->entry == &cursor->parent->entries[num_entries - 1])
+               return false;
+
+       cursor->pfn += 1ULL << shift;
+       cursor->pfn &= ~((1ULL << shift) - 1);
+       ++cursor->entry;
+       return true;
+}
+
+/**
+ * amdgpu_vm_pt_ancestor - go to parent node
+ *
+ * @cursor: current state
+ *
+ * Walk to the parent node of the current node.
+ * Returns:
+ * True if the walk was possible, false otherwise.
+ */
+static bool amdgpu_vm_pt_ancestor(struct amdgpu_vm_pt_cursor *cursor)
+{
+       if (!cursor->parent)
+               return false;
+
+       --cursor->level;
+       cursor->entry = cursor->parent;
+       cursor->parent = amdgpu_vm_pt_parent(cursor->parent);
+       return true;
+}
+
+/**
+ * amdgpu_vm_pt_next - get next PD/PT in hieratchy
+ *
+ * @adev: amdgpu_device pointer
+ * @cursor: current state
+ *
+ * Walk the PD/PT tree to the next node.
+ */
+static void amdgpu_vm_pt_next(struct amdgpu_device *adev,
+                             struct amdgpu_vm_pt_cursor *cursor)
+{
+       /* First try a newborn child */
+       if (amdgpu_vm_pt_descendant(adev, cursor))
+               return;
+
+       /* If that didn't worked try to find a sibling */
+       while (!amdgpu_vm_pt_sibling(adev, cursor)) {
+               /* No sibling, go to our parents and grandparents */
+               if (!amdgpu_vm_pt_ancestor(cursor)) {
+                       cursor->pfn = ~0ll;
+                       return;
+               }
+       }
+}
+
+/**
+ * amdgpu_vm_pt_first_leaf - get first leaf PD/PT
+ *
+ * @adev: amdgpu_device pointer
+ * @vm: amdgpu_vm structure
+ * @start: start addr of the walk
+ * @cursor: state to initialize
+ *
+ * Start a walk and go directly to the leaf node.
+ */
+static void amdgpu_vm_pt_first_leaf(struct amdgpu_device *adev,
+                                   struct amdgpu_vm *vm, uint64_t start,
+                                   struct amdgpu_vm_pt_cursor *cursor)
+{
+       amdgpu_vm_pt_start(adev, vm, start, cursor);
+       while (amdgpu_vm_pt_descendant(adev, cursor));
+}
+
+/**
+ * amdgpu_vm_pt_next_leaf - get next leaf PD/PT
+ *
+ * @adev: amdgpu_device pointer
+ * @cursor: current state
+ *
+ * Walk the PD/PT tree to the next leaf node.
+ */
+static void amdgpu_vm_pt_next_leaf(struct amdgpu_device *adev,
+                                  struct amdgpu_vm_pt_cursor *cursor)
+{
+       amdgpu_vm_pt_next(adev, cursor);
+       while (amdgpu_vm_pt_descendant(adev, cursor));
+}
+
+/**
+ * for_each_amdgpu_vm_pt_leaf - walk over all leaf PDs/PTs in the hierarchy
+ */
+#define for_each_amdgpu_vm_pt_leaf(adev, vm, start, end, cursor)               \
+       for (amdgpu_vm_pt_first_leaf((adev), (vm), (start), &(cursor));         \
+            (cursor).pfn <= end; amdgpu_vm_pt_next_leaf((adev), &(cursor)))
+
+/**
+ * amdgpu_vm_pt_first_dfs - start a deep first search
+ *
+ * @adev: amdgpu_device structure
+ * @vm: amdgpu_vm structure
+ * @cursor: state to initialize
+ *
+ * Starts a deep first traversal of the PD/PT tree.
+ */
+static void amdgpu_vm_pt_first_dfs(struct amdgpu_device *adev,
+                                  struct amdgpu_vm *vm,
+                                  struct amdgpu_vm_pt_cursor *cursor)
+{
+       amdgpu_vm_pt_start(adev, vm, 0, cursor);
+       while (amdgpu_vm_pt_descendant(adev, cursor));
+}
+
+/**
+ * amdgpu_vm_pt_next_dfs - get the next node for a deep first search
+ *
+ * @adev: amdgpu_device structure
+ * @cursor: current state
+ *
+ * Move the cursor to the next node in a deep first search.
+ */
+static void amdgpu_vm_pt_next_dfs(struct amdgpu_device *adev,
+                                 struct amdgpu_vm_pt_cursor *cursor)
+{
+       if (!cursor->entry)
+               return;
+
+       if (!cursor->parent)
+               cursor->entry = NULL;
+       else if (amdgpu_vm_pt_sibling(adev, cursor))
+               while (amdgpu_vm_pt_descendant(adev, cursor));
+       else
+               amdgpu_vm_pt_ancestor(cursor);
+}
+
+/**
+ * for_each_amdgpu_vm_pt_dfs_safe - safe deep first search of all PDs/PTs
+ */
+#define for_each_amdgpu_vm_pt_dfs_safe(adev, vm, cursor, entry)                        \
+       for (amdgpu_vm_pt_first_dfs((adev), (vm), &(cursor)),                   \
+            (entry) = (cursor).entry, amdgpu_vm_pt_next_dfs((adev), &(cursor));\
+            (entry); (entry) = (cursor).entry,                                 \
+            amdgpu_vm_pt_next_dfs((adev), &(cursor)))
+
 /**
  * amdgpu_vm_get_pd_bo - add the VM PD to a validation list
  *