Re: [PATCH v4 1/4] [media] davinci: vpif_capture: don't lock over s_stream

[Laurent Pinchart <laurent.pinchart@ideasonboard.com>]

Hi Kevin,

On Tuesday 06 Dec 2016 08:49:38 Kevin Hilman wrote:
> Laurent Pinchart writes:
> > On Tuesday 29 Nov 2016 15:57:09 Kevin Hilman wrote:
> >> Video capture subdevs may be over I2C and may sleep during xfer, so we
> >> cannot do IRQ-disabled locking when calling the subdev.
> >> 
> >> Signed-off-by: Kevin Hilman <khilman@baylibre.com>
> >> ---
> >>  drivers/media/platform/davinci/vpif_capture.c | 3 +++
> >>  1 file changed, 3 insertions(+)
> >> 
> >> diff --git a/drivers/media/platform/davinci/vpif_capture.c
> >> b/drivers/media/platform/davinci/vpif_capture.c index
> >> 5104cc0ee40e..9f8f41c0f251 100644
> >> --- a/drivers/media/platform/davinci/vpif_capture.c
> >> +++ b/drivers/media/platform/davinci/vpif_capture.c
> >> @@ -193,7 +193,10 @@ static int vpif_start_streaming(struct vb2_queue
> >> *vq, unsigned int count)
> >>  		}
> >>  	}
> >> 
> >> +	spin_unlock_irqrestore(&common->irqlock, flags);
> >>  	ret = v4l2_subdev_call(ch->sd, video, s_stream, 1);
> >> +	spin_lock_irqsave(&common->irqlock, flags);
> > 
> > I always get anxious when I see a spinlock being released randomly with an
> > operation in the middle of a protected section. Looking at the code it
> > looks like the spinlock is abused here. irqlock should only protect the
> > dma_queue and should thus only be taken around the following code:
> > 
> > spin_lock_irqsave(&common->irqlock, flags);
> > /* Get the next frame from the buffer queue */
> > common->cur_frm = common->next_frm = list_entry(common->dma_queue.next,
> >                             struct vpif_cap_buffer, list);
> > 
> > /* Remove buffer from the buffer queue */
> > list_del(&common->cur_frm->list);
> > spin_unlock_irqrestore(&common->irqlock, flags);
> 
> Yes, that looks correct.  Will update.
> 
> > The code that is currently protected by the lock in the start and stop
> > streaming functions should be protected by a mutex instead.
> 
> I tried taking the mutex here, but lockdep pointed out a deadlock.  I
> may not be fully understanding the V4L2 internals here, but it seems
> that the ioctl is already taking a mutex, so taking it again in
> start/stop streaming is a deadlock.  Unless you think the locking should
> be nested here, it seems to me that the mutex isn't needed.

The V4L2 core can lock all ioctls using struct video_device::lock. For buffer-
related ioctls, it can optionally use a separate lock from struct 
vb2_queue::lock. See v4l2_ioctl_get_lock() in drivers/media/v4l2-core/v4l2-
ioctl.c.

The vpif-capture driver sets both the video_device and vb2_queue locks to the 
same lock (which would have the same effect as leaving the vb2_queue lock 
NULL). All ioctls are thus serialized. You would only need to handle locking 
in start_streaming and stop_streaming manually if you didn't rely on the core 
serializing the ioctls.

-- 
Regards,

Laurent Pinchart