Re: [PATCH net-next v3 5/9] device: add device_change_owner()

On Thu, Feb 20, 2020 at 12:25:13PM +0100, Greg Kroah-Hartman wrote:

On Tue, Feb 18, 2020 at 05:29:39PM +0100, Christian Brauner wrote:

quoted

Add a helper to change the owner of a device's sysfs entries. This
needs to happen when the ownership of a device is changed, e.g. when
moving network devices between network namespaces.
This function will be used to correctly account for ownership changes,
e.g. when moving network devices between network namespaces.

Signed-off-by: Christian Brauner <redacted>
---
/* v2 */
unchanged

/* v3 */
-  Greg Kroah-Hartman [off-list ref]:
   - Add explicit uid/gid parameters.
---
 drivers/base/core.c    | 80 ++++++++++++++++++++++++++++++++++++++++++
 include/linux/device.h |  1 +
 2 files changed, 81 insertions(+)

diff --git a/drivers/base/core.c b/drivers/base/core.c
index 42a672456432..ec0d5e8cfd0f 100644
--- a/drivers/base/core.c
+++ b/drivers/base/core.c

@@ -3458,6 +3458,86 @@ int device_move(struct device *dev, struct device *new_parent,
 }
 EXPORT_SYMBOL_GPL(device_move);
 
+static int device_attrs_change_owner(struct device *dev, kuid_t kuid,
+				     kgid_t kgid)
+{
+	struct kobject *kobj = &dev->kobj;
+	struct class *class = dev->class;
+	const struct device_type *type = dev->type;
+	int error;
+
+	if (class) {
+		error = sysfs_groups_change_owner(kobj, class->dev_groups, kuid,
+						  kgid);
+		if (error)
+			return error;
+	}
+
+	if (type) {
+		error = sysfs_groups_change_owner(kobj, type->groups, kuid,
+						  kgid);
+		if (error)
+			return error;
+	}
+
+	error = sysfs_groups_change_owner(kobj, dev->groups, kuid, kgid);
+	if (error)
+		return error;
+
+	if (device_supports_offline(dev) && !dev->offline_disabled) {
+		error = sysfs_file_change_owner_by_name(
+			kobj, dev_attr_online.attr.name, kuid, kgid);
+		if (error)
+			return error;
+	}
+
+	return 0;
+}
+
+/**
+ * device_change_owner - change the owner of an existing device.

The "owner" and what else gets changed here?  Please document this
better.

quoted

+ * @dev: device.
+ * @kuid: new owner's kuid
+ * @kgid: new owner's kgid
+ */
+int device_change_owner(struct device *dev, kuid_t kuid, kgid_t kgid)
+{
+	int error;
+	struct kobject *kobj = &dev->kobj;
+
+	dev = get_device(dev);
+	if (!dev)
+		return -EINVAL;
+
+	error = sysfs_change_owner(kobj, kuid, kgid);

the kobject of the device is changed, good.

quoted

+	if (error)
+		goto out;
+
+	error = sysfs_file_change_owner_by_name(kobj, dev_attr_uevent.attr.name,
+						kuid, kgid);

Why call out the uevent file explicitly here?

This again, mirrors the creation of a kobject in sysfs. The uevent file
is created separately and thus should be chowned separately.

quoted

+	if (error)
+		goto out;
+
+	error = device_attrs_change_owner(dev, kuid, kgid);
+	if (error)
+		goto out;

Doesn't this also change the uevent file?

No, not as far as I can tell. The uevent file is created in an extra
step when the kobject/sysfs entries are created.

quoted

+
+#ifdef CONFIG_BLOCK
+	if (sysfs_deprecated && dev->class == &block_class)
+		goto out;
+#endif

Ugh, we still need this?

Yeah, apparently. It's what I gather from how a device is added.

quoted

+
+	error = sysfs_link_change_owner(&dev->class->p->subsys.kobj, &dev->kobj,
+					dev_name(dev), kuid, kgid);

Now what is this changing?

So, this changed the ownership of the class link for the device to match
the directory entry for that device, so e.g. given a network device
symlink (or any other type) that points to the actual directory entry
for that device:

/sys/class/net/my-dev -> ../../devices/virtual/net/my-dev

it makes my-dev show the same permissions as the directory my-dev has.
If we don't do this this will look weird, because the symlink will show
different permissions than the target it is pointoing to.

Again, more documentation please as to exactly what is being changed in
this function is needed.

Sure!