On Monday, September 05, 2011 12:13:26 Hans Verkuil wrote:
> On Monday, September 05, 2011 11:59:03 Laurent Pinchart wrote:
> > Hi Hans,
> >
> > On Friday 02 September 2011 09:29:08 Sitsofe Wheeler wrote:
> > > On Fri, Sep 02, 2011 at 01:35:49PM +0800, Dave Young wrote:
> > > > On Fri, Sep 2, 2011 at 12:59 PM, Dave Young wrote:
> > > > > On Fri, Sep 2, 2011 at 3:10 AM, Sitsofe Wheeler wrote:
> > > > >> On Thu, Sep 01, 2011 at 05:02:51PM +0800, Dave Young wrote:
> > > > >>> On Tue, Aug 30, 2011 at 4:48 AM, Sitsofe Wheeler wrote:
> > > > >>> > I managed to produce an oops in 3.1.0-rc3-00270-g7a54f5e by
> > > > >>> > unplugging a
> > > > >>>
> > > > >>> > USB webcam. See below:
> > > > >>> Could you try the attached patch?
> > > > >>
> > > > >> This patch fixed the oops but extending the sequence (enable camera,
> > > > >> start cheese, disable camera, watch cheese pause, enable camera, quit
> > > > >> cheese, start cheese) causes the following "poison overwritten"
> > > > >> warning
> > > > >
> > > > >> to appear:
> > > > > It seems another bug, I can reproduce this as well.
> > > > >
> > > > > uvc_device is freed in uvc_delete,
> > > > >
> > > > > struct v4l2_device vdev is the member of struct uvc_device, so vdev is
> > > > > also freed. Later v4l2_device operations on vdev will overwrite the
> > > > > poison memory area.
> > > >
> > > > Please try attached patch on top of previous one, in this patch I
> > > > move v4l2_device_put after vdev->release in function
> > > > v4l2_device_release
> > > >
> > > > Not sure if this is a right fix, comments?
> >
> > (inlining the patch's contents)
> >
> > > > diff --git a/drivers/media/video/v4l2-dev.c b/drivers/media/video/v4l2-
> > > > dev.c
> > > > index 98cee19..541dba3 100644
> > > > --- a/drivers/media/video/v4l2-dev.c
> > > > +++ b/drivers/media/video/v4l2-dev.c
> > > > @@ -172,13 +172,14 @@ static void v4l2_device_release(struct device *cd)
> > > > media_device_unregister_entity(&vdev->entity);
> > > > #endif
> > > >
> > > > + /* Decrease v4l2_device refcount */
> > > > + if (vdev->v4l2_dev)
> > > > + v4l2_device_put(vdev->v4l2_dev);
> > > > +
> > > > /* Release video_device and perform other
> > > > cleanups as needed. */
> > > > vdev->release(vdev);
> > > >
> > > > - /* Decrease v4l2_device refcount */
> > > > - if (vdev->v4l2_dev)
> > > > - v4l2_device_put(vdev->v4l2_dev);
> > > > }
> > > >
> > > > static struct class video_class = {
> >
> > v4l2_device_put() got introduced in commit
> > bedf8bcf6b4f90a6e31add3721a2e71877289381 ("v4l2-device: add kref and a release
> > function"). If I understand its purpose correctly, drivers that use a
> > v4l2_device instance should use the v4l2_device release callback to release
> > device structures instead of counting video_device release callbacks manually.
> > In that case I think the v4l2-dev.c code is correct, and all drivers that use
> > v4l2_device should be fixed.
> >
> > The above patch fixes the problem in a central location, but seems to defeat
> > the original purpose of v4l2_device_get/put().
> >
> > Hans, could you please comment on that ?
>
> The original order is correct, but what I missed is that for drivers that release
> (free) everything in the videodev release callback the v4l2_device struct is
> also freed and v4l2_device_put will fail.
>
> To fix this, add this code just before the vdev->release call:
>
> /* Do not call v4l2_device_put if there is no release callback set. */
> if (v4l2_dev->release == NULL)
> v4l2_dev = NULL;
>
> If there is no release callback, then the refcounting is pointless anyway.
>
> This should work.
Note that in the long run using the v4l2_device release callback instead of the
videodev release is better. But it's a lot of work to convert everything so that's
long term. I'm quite surprised BTW that this bug wasn't found much earlier.
Regards,
Hans
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
