Skip to contentSkip to conversation

media: Revert 2a934fdb01db ("media: v4l2-dev: fix error handling in __video_register_device()")

PATCH In review Hans Verkuil 2 participants 4 messages Jul 17 – Jul 27

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover media: Revert 2a934fdb01db ("media: v4l2-dev: fix error handling in __video_register_device()")

That commit exchanged a memory leak with a double-free scenario, and it is not something that is easily fixable. Revert it, since it is still better to leak memory than get into a double-free.

See this email discussion for more info:

https://lore.kernel.org/linux-media/[email protected]/

I tried to come up with a reliable method to handle this cleanly in the v4l2 core, but the only way to do this properly would be to split video_register_device into two parts: an init and a registration part. And then convert all drivers to use the new functions.

That's a very big job, and so for now just revert the patch and add comments explaining why it is reverted.

Regards,

Hans

Hans Verkuil (2):
Revert "media: v4l2-dev: fix error handling in

    __video_register_device()"

media: v4l2-core: v4l2-dev: add comments on device_register fail.

 drivers/media/v4l2-core/v4l2-dev.c | 28 +++++++++++++++++++++-------
 1 file changed, 21 insertions(+), 7 deletions(-)
Signature · 1 line
2.53.0
#1 lore raw

PATCH 1/2 Revert "media: v4l2-dev: fix error handling in __video_register_device()"

This reverts commit 2a934fdb01db6458288fc9386d3d8ceba6dd551a.

The intentions of that patch were good, but it doesn't work.

The idea is that if device_register fails, you have to do a put_device to let the ref counter release resources.

However, the V4L2 API says that if video_register_device() fails, then you have to call video_device_release(), which kfree()s the video_device struct.

But the put_device() will already have freed the struct, so you end up in a double-free scenario.

There is not really a good way of fixing this without breaking video_register_device() into two parts, one that initializes everything, and one that does the actual device_register, and then converting all V4L2 drivers to this new model.

That is a massive job, and it is very unlikely that device_register will fail.

So rather than ending up in a double-free scenario, just revert this patch, and in that case we'll have a small memory leak. Which is a lot more robust.

---
 drivers/media/v4l2-core/v4l2-dev.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
drivers/media/v4l2-core/v4l2-dev.c +7 -7
index 5516b2bbb08f..d750bf10febe 100644
@@ -1071,25 +1071,25 @@ int __video_register_device(struct video_device *vdev,
 	vdev->dev.class = &video_class;
 	vdev->dev.devt = MKDEV(VIDEO_MAJOR, vdev->minor);
 	vdev->dev.parent = vdev->dev_parent;
-	vdev->dev.release = v4l2_device_release;
 	dev_set_name(&vdev->dev, "%s%d", name_base, vdev->num);
-
-	/* Increase v4l2_device refcount */
-	v4l2_device_get(vdev->v4l2_dev);
-
 	mutex_lock(&videodev_lock);
 	ret = device_register(&vdev->dev);
 	if (ret < 0) {
 		mutex_unlock(&videodev_lock);
 		pr_err("%s: device_register failed\n", __func__);
-		put_device(&vdev->dev);
-		return ret;
+		goto cleanup;
 	}
+	/* Register the release callback that will be called when the last
+	   reference to the device goes away. */
+	vdev->dev.release = v4l2_device_release;
 
 	if (nr != -1 && nr != vdev->num && warn_if_nr_in_use)
 		pr_warn("%s: requested %s%d, got %s\n", __func__,
 			name_base, nr, video_device_node_name(vdev));
 
+	/* Increase v4l2_device refcount */
+	v4l2_device_get(vdev->v4l2_dev);
+
 	/* Part 5: Register the entity. */
 	ret = video_register_media_controller(vdev);
 
Signature · 1 line
2.53.0
#2 lore raw

PATCH 2/2 media: v4l2-core: v4l2-dev: add comments on device_register fail.

In reply to Hans Verkuil

If device_register fails, then we are supposed to call put_device. Explain why we do not do that.

---
 drivers/media/v4l2-core/v4l2-dev.c | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
drivers/media/v4l2-core/v4l2-dev.c +14 -0
index d750bf10febe..fd267fb74905 100644
@@ -1075,6 +1075,20 @@ int __video_register_device(struct video_device *vdev,
 	mutex_lock(&videodev_lock);
 	ret = device_register(&vdev->dev);
 	if (ret < 0) {
+		/*
+		 * We should do a put_device() here, but the problem is that
+		 * the V4L2 API expects drivers to call video_device_release()
+		 * on error, and so both put_device() and video_device_release
+		 * would kfree vdev.
+		 *
+		 * The proper solution would be to split this function into
+		 * two parts: initialization and registration, and then rework
+		 * all drivers.
+		 *
+		 * Until then just skip the put_device and free everything.
+		 * This will result in a small memory leak, which is better
+		 * than a double-free.
+		 */
 		mutex_unlock(&videodev_lock);
 		pr_err("%s: device_register failed\n", __func__);
 		goto cleanup;
Signature · 1 line
2.53.0
Reviewed-by on the v1 series
#3 lore raw
In reply to Hans Verkuil
Quoted 4 lines · Hans Verkuil

That commit exchanged a memory leak with a double-free scenario, and it is not something that is easily fixable. Revert it, since it is still better to leak memory than get into a double-free.

Quoted 8 lines

See this email discussion for more info:

https://lore.kernel.org/linux-media/[email protected]/

I tried to come up with a reliable method to handle this cleanly in the v4l2 core, but the only way to do this properly would be to split video_register_device into two parts: an init and a registration part. And then convert all drivers to use the new functions.

We should do that.

Quoted 2 lines

That's a very big job, and so for now just revert the patch and add comments explaining why it is reverted.

Agreed. For the series,

Quoted 7 lines

Hans Verkuil (2):
Revert "media: v4l2-dev: fix error handling in

    __video_register_device()"

media: v4l2-core: v4l2-dev: add comments on device_register fail.

 drivers/media/v4l2-core/v4l2-dev.c | 28 +++++++++++++++++++++-------
 1 file changed, 21 insertions(+), 7 deletions(-)
Signature · 3 lines
Regards,

Laurent Pinchart