On Wed, Mar 21, 2012 at 01:44:36AM +0800, Stephen Warren wrote:

During pinctrl_get(), if the client device has a device tree node, look
for the common pinctrl properties there. If found, parse the referenced
device tree nodes, with the help of the pinctrl drivers, and generate
mapping table entries from them.

During pinctrl_put(), free any results of device tree parsing.

Signed-off-by: Stephen Warren <redacted>
---
v2: Place most code into new file devicetree.c

Much clearer.:)

+ * struct pinctrl_dt_map - mapping table chunk parsed from device tree
+ * @node: list node for struct pinctrl's @dt_maps field
+ * @pctldev: the pin controller that allocated this struct, and will free it
+ * @maps: the mapping table entries
+ */
+struct pinctrl_dt_map {
+	struct list_head node;
+	struct pinctrl_dev *pctldev;
+	struct pinctrl_map *map;
+	unsigned num_maps;
+};
+
+static void dt_free_map(struct pinctrl_dev *pctldev,
+		     struct pinctrl_map *map, unsigned num_maps)
+{
+	if (pctldev) {
+		struct pinctrl_ops *ops = pctldev->desc->pctlops;
+		ops->dt_free_map(pctldev, map, num_maps);
+	} else {

I remember for hog on functions the pctldev becomes pinctrl devices itself,
so in which case pctldev can be NULL?

+static struct pinctrl_dev *find_pinctrl_by_of_node(struct device_node *np)
+{
+	struct pinctrl_dev *pctldev;
+
+	list_for_each_entry(pctldev, &pinctrldev_list, node)
+		if (pctldev->dev->of_node == np)
+			return pctldev;
+
+	return NULL;
+}
+
+static int dt_to_map_one_config(struct pinctrl *p, const char *statename,
+				struct device_node *np_config)
+{
+	struct device_node *np_pctldev;
+	struct pinctrl_dev *pctldev;
+	struct pinctrl_ops *ops;
+	int ret;
+	struct pinctrl_map *map;
+	unsigned num_maps;
+
+	/* Find the pin controller containing np_config */
+	np_pctldev = of_node_get(np_config);

It seems the np_config node is already got when call of_find_node_by_phandle.
So do we still need this line?

+int pinctrl_dt_to_map(struct pinctrl *p)
+{
+	struct device_node *np = p->dev->of_node;
+	int state, ret;
+	char *propname;
+	struct property *prop;
+	const char *statename;
+	const __be32 *list;
+	int size, config;
+	phandle phandle;
+	struct device_node *np_config;
+	struct pinctrl_dt_map *dt_map;
+

Add NULL np checking?

+	/* We may store pointers to property names within the node */
+	of_node_get(np);
+
+	/* For each defined state ID */
+	for (state = 0; ; state++) {
+		/* Retrieve the pinctrl-* property */
+		propname = kasprintf(GFP_KERNEL, "pinctrl-%d", state);
+		prop = of_find_property(np, propname, &size);
+		kfree(propname);
+		if (!prop)
+			break;
+		list = prop->value;
+		size /= sizeof(*list);
+
+		/* Determine whether pinctrl-names property names the state */
+		ret = of_property_read_string_index(np, "pinctrl-names",
+						    state, &statename);
+		/*
+		 * If not, statename is just the integer state ID. But rather
+		 * than dynamically allocate it and have to free it later,
+		 * just point part way into the property name for the string.
+		 */
+		if (ret < 0) {
+			/* strlen("pinctrl-") == 8 */
+			if (strlen(prop->name) < 8) {

Do we really need this extra checking?
It seems the prop->name is the "pinctrl-%d" you searched above, so the
strlen(prop->name) must not < 8, right?

+				dev_err(p->dev, "prop %s inconsistent length\n",
+					prop->name);
+				ret = -EINVAL;
+				goto err;
+			}
+			statename = prop->name + 8;
From this code, it seems actually we provide user the option by chance to define

state as pinctrl-syspend which is out of our binding doc.

+		}
+
+		/* For every referenced pin configuration node in it */
+		for (config = 0; config < size; config++) {
+			phandle = be32_to_cpup(list++);
+
+			/* Look up the pin configuration node */
+			np_config = of_find_node_by_phandle(phandle);

One option is using of_parse_phandle, then we do not need calculate
the phandle offset by ourselves.
Like:
np_config = of_parse_phandle(propname , config);

+			if (!np_config) {
+				dev_err(p->dev,
+					"prop %s index %i invalid phandle\n",
+					prop->name, config);
+				ret = -EINVAL;
+				goto err;
+			}
+
+			/* Parse the node */
+			ret = dt_to_map_one_config(p, statename, np_config);
+			of_node_put(np_config);
+			if (ret < 0)
+				goto err;
+		}
+
+		/* No entries in DT? Generate a dummy state table entry */
+		if (!size) {
+			ret = dt_remember_dummy_state(p, statename);
+			if (ret < 0)
+				goto err;
+		}
+	}
+
+	list_for_each_entry(dt_map, &p->dt_maps, node) {
+		ret = pinctrl_register_map(dt_map->map, dt_map->num_maps,
+					   false, true);
+		if (ret < 0)
+			goto err;
+	}

What's main purpose of differing the map registration and introduce a
intermediate pinctrl_dt_map(dt_remember_or_free_map)?
What about directly register maps once it's parsed?

Regards
Dong Aisheng