dm btree: introduce cursor api

This uses prefetching to speed up iteration through a btree.

Signed-off-by: Joe Thornber <[e-mail]>
Signed-off-by: Mike Snitzer <[e-mail]>

drivers/md/persistent-data/dm-btree.c

···
 	return walk_node(info, root, fn, context);
 }
 EXPORT_SYMBOL_GPL(dm_btree_walk);
+
+/*----------------------------------------------------------------*/
+
+static void prefetch_values(struct dm_btree_cursor *c)
+{
+	unsigned i, nr;
+	__le64 value_le;
+	struct cursor_node *n = c->nodes + c->depth - 1;
+	struct btree_node *bn = dm_block_data(n->b);
+	struct dm_block_manager *bm = dm_tm_get_bm(c->info->tm);
+
+	BUG_ON(c->info->value_type.size != sizeof(value_le));
+
+	nr = le32_to_cpu(bn->header.nr_entries);
+	for (i = 0; i < nr; i++) {
+		memcpy(&value_le, value_ptr(bn, i), sizeof(value_le));
+		dm_bm_prefetch(bm, le64_to_cpu(value_le));
+	}
+}
+
+static bool leaf_node(struct dm_btree_cursor *c)
+{
+	struct cursor_node *n = c->nodes + c->depth - 1;
+	struct btree_node *bn = dm_block_data(n->b);
+
+	return le32_to_cpu(bn->header.flags) & LEAF_NODE;
+}
+
+static int push_node(struct dm_btree_cursor *c, dm_block_t b)
+{
+	int r;
+	struct cursor_node *n = c->nodes + c->depth;
+
+	if (c->depth >= DM_BTREE_CURSOR_MAX_DEPTH - 1) {
+		DMERR("couldn't push cursor node, stack depth too high");
+		return -EINVAL;
+	}
+
+	r = bn_read_lock(c->info, b, &n->b);
+	if (r)
+		return r;
+
+	n->index = 0;
+	c->depth++;
+
+	if (c->prefetch_leaves || !leaf_node(c))
+		prefetch_values(c);
+
+	return 0;
+}
+
+static void pop_node(struct dm_btree_cursor *c)
+{
+	c->depth--;
+	unlock_block(c->info, c->nodes[c->depth].b);
+}
+
+static int inc_or_backtrack(struct dm_btree_cursor *c)
+{
+	struct cursor_node *n;
+	struct btree_node *bn;
+
+	for (;;) {
+		if (!c->depth)
+			return -ENODATA;
+
+		n = c->nodes + c->depth - 1;
+		bn = dm_block_data(n->b);
+
+		n->index++;
+		if (n->index < le32_to_cpu(bn->header.nr_entries))
+			break;
+
+		pop_node(c);
+	}
+
+	return 0;
+}
+
+static int find_leaf(struct dm_btree_cursor *c)
+{
+	int r = 0;
+	struct cursor_node *n;
+	struct btree_node *bn;
+	__le64 value_le;
+
+	for (;;) {
+		n = c->nodes + c->depth - 1;
+		bn = dm_block_data(n->b);
+
+		if (le32_to_cpu(bn->header.flags) & LEAF_NODE)
+			break;
+
+		memcpy(&value_le, value_ptr(bn, n->index), sizeof(value_le));
+		r = push_node(c, le64_to_cpu(value_le));
+		if (r) {
+			DMERR("push_node failed");
+			break;
+		}
+	}
+
+	if (!r && (le32_to_cpu(bn->header.nr_entries) == 0))
+		return -ENODATA;
+
+	return r;
+}
+
+int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
+			  bool prefetch_leaves, struct dm_btree_cursor *c)
+{
+	int r;
+
+	c->info = info;
+	c->root = root;
+	c->depth = 0;
+	c->prefetch_leaves = prefetch_leaves;
+
+	r = push_node(c, root);
+	if (r)
+		return r;
+
+	return find_leaf(c);
+}
+EXPORT_SYMBOL_GPL(dm_btree_cursor_begin);
+
+void dm_btree_cursor_end(struct dm_btree_cursor *c)
+{
+	while (c->depth)
+		pop_node(c);
+}
+EXPORT_SYMBOL_GPL(dm_btree_cursor_end);
+
+int dm_btree_cursor_next(struct dm_btree_cursor *c)
+{
+	int r = inc_or_backtrack(c);
+	if (!r) {
+		r = find_leaf(c);
+		if (r)
+			DMERR("find_leaf failed");
+	}
+
+	return r;
+}
+EXPORT_SYMBOL_GPL(dm_btree_cursor_next);
+
+int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le)
+{
+	if (c->depth) {
+		struct cursor_node *n = c->nodes + c->depth - 1;
+		struct btree_node *bn = dm_block_data(n->b);
+
+		if (le32_to_cpu(bn->header.flags) & INTERNAL_NODE)
+			return -EINVAL;
+
+		*key = le64_to_cpu(*key_ptr(bn, n->index));
+		memcpy(value_le, value_ptr(bn, n->index), c->info->value_type.size);
+		return 0;
+
+	} else
+		return -ENODATA;
+}
+EXPORT_SYMBOL_GPL(dm_btree_cursor_get_value);

drivers/md/persistent-data/dm-btree.h

···
 		  int (*fn)(void *context, uint64_t *keys, void *leaf),
 		  void *context);
 
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Cursor API.  This does not follow the rolling lock convention.  Since we
+ * know the order that values are required we can issue prefetches to speed
+ * up iteration.  Use on a single level btree only.
+ */
+#define DM_BTREE_CURSOR_MAX_DEPTH 16
+
+struct cursor_node {
+	struct dm_block *b;
+	unsigned index;
+};
+
+struct dm_btree_cursor {
+	struct dm_btree_info *info;
+	dm_block_t root;
+
+	bool prefetch_leaves;
+	unsigned depth;
+	struct cursor_node nodes[DM_BTREE_CURSOR_MAX_DEPTH];
+};
+
+/*
+ * Creates a fresh cursor.  If prefetch_leaves is set then it is assumed
+ * the btree contains block indexes that will be prefetched.  The cursor is
+ * quite large, so you probably don't want to put it on the stack.
+ */
+int dm_btree_cursor_begin(struct dm_btree_info *info, dm_block_t root,
+			  bool prefetch_leaves, struct dm_btree_cursor *c);
+void dm_btree_cursor_end(struct dm_btree_cursor *c);
+int dm_btree_cursor_next(struct dm_btree_cursor *c);
+int dm_btree_cursor_get_value(struct dm_btree_cursor *c, uint64_t *key, void *value_le);
+
 #endif	/* _LINUX_DM_BTREE_H */