Re: Linux 2.6.26-rc9 circular lock with uvcvideo on resume from hibernation

[Roland Dreier <roland@digitalvampire.org>]

 > uvc_disconnect()                      | uvc_v4l2_open()
 > ...                                   |
 > mutex_lock(&uvc_driver.open_mutex);   |
 > dev->state |= UVC_DEV_DISCONNECTED;   |
 > mutex_unlock(&uvc_driver.open_mutex); |
 >                                       |
 >                                       | mutex_lock(&uvc_driver.open_mutex);
 >                                       | vdev = video_devdata(file);
 >                                       | video = video_get_drvdata(vdev);
 >                                       |
 > kref_put(&dev->kref, uvc_delete);     |
 >                                       |
 >                                       | if (video->dev->state...) 
 > 
 > kref_put() in uvc_disconnect() will call uvc_delete(), which will in turn free 
 > the video structure. uvc_v4l2_open() will then dereference freed memory when 
 > testing the device state.

I don't believe this is correct.  I tried to explain it in my
changelog by saying "uvc_delete() does uvc_unregister_video() (and
hence video_unregister_device(), which is synchronized with
videodev_lock) as its first thing, so there is no risk of
use-after-free in uvc_v4l2_open()."

In other words, the first thing uvc_delete() does is call
uvc_unregister_video(), which will video_unregister_device().  Since
this needs to take videodev_lock, it will wait until uvc_v4l2_open()
returns (which it will do, since state is now UVC_DEV_DISCONNECTED).
So the video struct will not be freed until after uvc_v4l2_open()
returns.

As far as I can see there is no use-after-free.

 - R.
-- 
Roland Dreier  <roland@digitalvampire.org>  GPG Key: 1024D/E0EEFAC0
Fingerprint:     A89F B5E9 C185 F34D BD50  4009 37E2 25CC E0EE FAC0

 Sending >500KB attachments is forbidden by the Geneva Convention.
        Your country may be at risk if you fail to comply.