The uvcvideo driver's uvc_v4l2_open() method is called from videodev's
video_open() function, which means it is called with the videodev_lock
mutex held. uvc_v4l2_open() then takes uvc_driver.open_mutex to check
dev->state and avoid racing against a device disconnect, which means
that open_mutex must nest inside videodev_lock.
However uvc_disconnect() takes the open_mutex around setting
dev->state and also around putting its device reference. However, if
uvc_disconnect() ends up dropping the last reference, it will call
uvc_delete(), which calls into the videodev code to unregister its
device, and this will end up taking videodev_lock. This opens a
(unlikely in practice) window for an AB-BA deadlock and also causes a
lockdep warning because of the lock misordering.
Fortunately there is no apparent reason to hold open_mutex when doing
kref_put() in uvc_disconnect(): if uvc_v4l2_open() runs before the
state is set to UVC_DEV_DISCONNECTED, then it will take another
reference to the device and kref_put() won't call uvc_delete; if
uvc_v4l2_open() runs after the state is set, it will run before
uvc_delete(), see the state, and return immediately -- 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().
Bug diagnosed based on a lockdep warning reported by Romano Giannetti
<romano@dea.icai.upcomillas.es>.
Signed-off-by: Roland Dreier <roland@digitalvampire.org>
Signed-off-by: Laurent Pinchart <laurent.pinchart@skynet.be>
Signed-off-by: Mauro Carvalho Chehab <mchehab@infradead.org>
a9e2858561