Re: [PATCH v5 3/3] media: uvcvideo: Do not use usb_* functions after .disconnect

From: Laurent Pinchart
Date: Fri Mar 22 2024 - 10:19:20 EST

Hi Ricardo,

On Wed, Nov 22, 2023 at 03:59:10PM +0100, Ricardo Ribalda wrote:
> On Wed, 22 Nov 2023 at 14:33, Laurent Pinchart wrote:
> > On Wed, Nov 22, 2023 at 11:45:49AM +0000, Ricardo Ribalda wrote:
> > > usb drivers should not call to any I/O function after the
> > > .disconnect() callback has been triggered.
> > > https://www.kernel.org/doc/html/latest/driver-api/usb/callbacks.html#the-disconnect-callback
> > >
> > > If an application is receiving frames form a camera and the device is
> > > disconnected: the device will call close() after the usb .disconnect()
> > > callback has been called. The streamoff path will call usb_set_interface
> > > or usb_clear_halt, which is not allowed.
> > >
> > > This patch only solves the calls to close() *after* .disconnect() is
> > > being called.
> > >
> > > Trace:
> > > [ 1065.389723] drivers/media/usb/uvc/uvc_driver.c:2248 uvc_disconnect enter
> > > [ 1065.390160] drivers/media/usb/uvc/uvc_driver.c:2264 uvc_disconnect exit
> > > [ 1065.433956] drivers/media/usb/uvc/uvc_v4l2.c:659 uvc_v4l2_release enter
> > > [ 1065.433973] drivers/media/usb/uvc/uvc_video.c:2274 uvc_video_stop_streaming enter
> > > [ 1065.434560] drivers/media/usb/uvc/uvc_video.c:2285 uvc_video_stop_streaming exit
> > > [ 1065.435154] drivers/media/usb/uvc/uvc_v4l2.c:680 uvc_v4l2_release exit
> > > [ 1065.435188] drivers/media/usb/uvc/uvc_driver.c:2248 uvc_disconnect enter
> > >
> > > Signed-off-by: Ricardo Ribalda <ribalda@xxxxxxxxxxxx>
> > > ---
> > > drivers/media/usb/uvc/uvc_driver.c | 2 ++
> > > drivers/media/usb/uvc/uvc_video.c | 45 ++++++++++++++++++++++++--------------
> > > drivers/media/usb/uvc/uvcvideo.h | 2 ++
> > > 3 files changed, 32 insertions(+), 17 deletions(-)
> > >
> > > diff --git a/drivers/media/usb/uvc/uvc_driver.c b/drivers/media/usb/uvc/uvc_driver.c
> > > index d5dbf2644272..d78640d422f4 100644
> > > --- a/drivers/media/usb/uvc/uvc_driver.c
> > > +++ b/drivers/media/usb/uvc/uvc_driver.c
> > > @@ -2266,6 +2266,8 @@ static void uvc_disconnect(struct usb_interface *intf)
> > > return;
> > >
> > > uvc_unregister_video(dev);
> > > + /* Barrier needed to pair with uvc_video_stop_streaming(). */
> > > + smp_store_release(&dev->disconnected, true);
> >
> > I can't think this would be such a hot path that we really need barriers
> > in the driver.
>
> Using the same variable from two contexts without any sync feels weird.
>
> Your concern is that there will be a big penalty by using the
> barriers? This is only used in stop_streaming and the shutdown path.

What I meant is that lockless concurrency is much harder to get right
than locked concurrency. Unless there's an important reason not to use
lock (which would I think need to be related to performance, and I don't
see that being an issue here), I think using locks will be less
error-prone and more maintainable. That's the KISS approach to
concurrency (even if it doesn't directly address lockless concurrency, I
like the approach advocated by Sima in
https://blog.ffwll.ch/2022/07/locking-engineering.html).

> > > kref_put(&dev->ref, uvc_delete);
> > > }
> > >
> > > diff --git a/drivers/media/usb/uvc/uvc_video.c b/drivers/media/usb/uvc/uvc_video.c
> > > index 28dde08ec6c5..f5ef375088de 100644
> > > --- a/drivers/media/usb/uvc/uvc_video.c
> > > +++ b/drivers/media/usb/uvc/uvc_video.c
> > > @@ -2243,28 +2243,39 @@ int uvc_video_start_streaming(struct uvc_streaming *stream)
> > > return ret;
> > > }
> > >
> > > -void uvc_video_stop_streaming(struct uvc_streaming *stream)
> > > +static void uvc_video_halt(struct uvc_streaming *stream)
> > > {
> > > - uvc_video_stop_transfer(stream, 1);
> > > + unsigned int epnum;
> > > + unsigned int pipe;
> > > + unsigned int dir;
> > >
> > > if (stream->intf->num_altsetting > 1) {
> > > usb_set_interface(stream->dev->udev, stream->intfnum, 0);
> > > - } else {
> > > - /*
> > > - * UVC doesn't specify how to inform a bulk-based device
> > > - * when the video stream is stopped. Windows sends a
> > > - * CLEAR_FEATURE(HALT) request to the video streaming
> > > - * bulk endpoint, mimic the same behaviour.
> > > - */
> > > - unsigned int epnum = stream->header.bEndpointAddress
> > > - & USB_ENDPOINT_NUMBER_MASK;
> > > - unsigned int dir = stream->header.bEndpointAddress
> > > - & USB_ENDPOINT_DIR_MASK;
> > > - unsigned int pipe;
> > > -
> > > - pipe = usb_sndbulkpipe(stream->dev->udev, epnum) | dir;
> > > - usb_clear_halt(stream->dev->udev, pipe);
> > > + return;
> > > }
> > >
> > > + /*
> > > + * UVC doesn't specify how to inform a bulk-based device
> > > + * when the video stream is stopped. Windows sends a
> > > + * CLEAR_FEATURE(HALT) request to the video streaming
> > > + * bulk endpoint, mimic the same behaviour.
> > > + */
> > > + epnum = stream->header.bEndpointAddress & USB_ENDPOINT_NUMBER_MASK;
> > > + dir = stream->header.bEndpointAddress & USB_ENDPOINT_DIR_MASK;
> > > + pipe = usb_sndbulkpipe(stream->dev->udev, epnum) | dir;
> > > + usb_clear_halt(stream->dev->udev, pipe);
> > > +}
> > > +
> > > +void uvc_video_stop_streaming(struct uvc_streaming *stream)
> > > +{
> > > + uvc_video_stop_transfer(stream, 1);
> > > +
> > > + /*
> > > + * Barrier needed to pair with uvc_disconnect().
> > > + * We cannot call usb_* functions on a disconnected USB device.
> > > + */
> > > + if (!smp_load_acquire(&stream->dev->disconnected))
> > > + uvc_video_halt(stream);
> > > +
> > > uvc_video_clock_cleanup(stream);
> > > }
> > > diff --git a/drivers/media/usb/uvc/uvcvideo.h b/drivers/media/usb/uvc/uvcvideo.h
> > > index ba8f8c1f2c83..5b1a3643de05 100644
> > > --- a/drivers/media/usb/uvc/uvcvideo.h
> > > +++ b/drivers/media/usb/uvc/uvcvideo.h
> > > @@ -559,6 +559,8 @@ struct uvc_device {
> > > unsigned int users;
> > > atomic_t nmappings;
> > >
> > > + bool disconnected;
> > > +
> > > /* Video control interface */
> > > #ifdef CONFIG_MEDIA_CONTROLLER
> > > struct media_device mdev;
> > >

--
Regards,

Laurent Pinchart