[lkml]   [2008]   [Dec]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [BUG] cdev_put() race condition
On Tuesday 16 December 2008 21:22:48 Greg KH wrote:
> On Mon, Dec 08, 2008 at 09:56:26PM +0100, Hans Verkuil wrote:
> > Hi Greg,
> >
> > Laurent found a race condition in the uvc driver that we traced to the
> > way chrdev_open and cdev_put/get work.
> >
> > You need the following ingredients to reproduce it:
> >
> > 1) a hot-pluggable char device like an USB webcam.
> > 2) a manually created device node for such a webcam instead of relying
> > on udev.
> >
> > In order to easily force this situation you would also need to add a
> > delay to the char device's release() function. For webcams that would
> > be at the top of v4l2_chardev_release() in
> > drivers/media/video/v4l2-dev.c. But adding a delay to e.g. cdev_purge
> > would have the same effect.
> >
> > The sequence of events in the case of a webcam is as follows:
> >
> > 1) The USB device is removed, causing a disconnect.
> >
> > 2) The webcam driver unregisters the video device which in turn calls
> > cdev_del().
> >
> > 3) When the last application using the device is closed, the cdev is
> > released when the kref of the cdev's kobject goes to 0.
> >
> > 4) If the kref's release() call takes a while due to e.g. extra cleanup
> > in the case of a webcam, then another application can try to open the
> > video device. Note that this requires a device node created with mknod,
> > otherwise the device nodes would already have been removed by udev.
> >
> > 5) chrdev_open checks inode->i_cdev. If this is NULL (i.e. this device
> > node was never accessed before), then all is fine since kobj_lookup
> > will fail because cdev_del() has been called earlier. However, if this
> > device node was used earlier, then the else part is called:
> > cdev_get(p). This 'p' is the cdev that is being released. Since the
> > kref count is 0 you will get a WARN message from kref_get, but the code
> > continues on, the f_op->open will (hopefully) return more-or-less
> > gracefully with an error and the cdev_put at the end will cause the
> > refcount to go to 0 again, which results in a SECOND call to the kref's
> > release function!
> >
> > See this link for the original discussion on the v4l list containing
> > stack traces an a patch that you need if you want to (and can) test
> > this with the uvc driver:
> >
> >
> The second sentence in that message shows your problem here:
> To avoid the need of a reference count in every v4l2 driver,
> v4l2 moved to cdev which includes its own reference counting
> infrastructure based on kobject.
> cdev is not ment to handle the reference counting of any object outside
> of itself, and should never be embedded within anything. I've been
> thinking of taking the real "kobject" out of that structure for a long
> time now, incase someone did something foolish like this.
> Seems I was too late :(
> So, to solve this, just remove the reliance on struct cdev in your own
> structures, you don't want to do this for the very reason you have now
> found (and for others, like the fact that this isn't a "real" struct
> kobject in play here, just a fake one.)
> Ick, what a mess.

Sorry, but this makes no sense. First of all the race condition exists regardless of how v4l uses it. Other drivers using cdev with a hot-pluggable device in combination with a manually created device node should show the same problem. It's just that we found it with v4l because the release callback takes longer than usual, thus increasing the chances of hitting the race.

The core problem is simply that it is possible to call cdev_get while in cdev_put! That should never happen.

Secondly, why shouldn't struct cdev be embedded in anything? It's used in lots of drivers that way. I really don't see what's unusual or messy about v4l in that respect.



Hans Verkuil - video4linux developer - sponsored by TANDBERG

 \ /
  Last update: 2008-12-16 22:03    [W:0.060 / U:0.856 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site