Re: [PATCH v3] media: allegro: add missed checks in allegro_open()

[Michael Tretter <m.tretter@pengutronix.de>]

On Tue, 21 Jan 2020 19:59:46 +0800, Chuhong Yuan wrote:
> On Tue, Jan 21, 2020 at 4:23 PM Michael Tretter
> <m.tretter@pengutronix.de> wrote:
> >
> > On Mon, 13 Jan 2020 13:59:51 +0800, Chuhong Yuan wrote:  
> > > allegro_open() misses checks for v4l2_m2m_ctx_init() and results of
> > > v4l2_ctrl_new* calls.
> > > Add checks and error handlers to fix the problems.
> > >
> > > Signed-off-by: Chuhong Yuan <hslester96@gmail.com>
> > > ---
> > > Changes in v3:
> > >   - Make code cleaner.
> > >   - Add a check for handler->error.
> > >
> > >  .../staging/media/allegro-dvt/allegro-core.c  | 24 +++++++++++++++----
> > >  1 file changed, 20 insertions(+), 4 deletions(-)
> > >
> > > diff --git a/drivers/staging/media/allegro-dvt/allegro-core.c b/drivers/staging/media/allegro-dvt/allegro-core.c
> > > index 6f0cd0784786..e86001e42963 100644
> > > --- a/drivers/staging/media/allegro-dvt/allegro-core.c
> > > +++ b/drivers/staging/media/allegro-dvt/allegro-core.c
> > > @@ -2270,15 +2270,12 @@ static int allegro_open(struct file *file)
> > >       struct allegro_channel *channel = NULL;
> > >       struct v4l2_ctrl_handler *handler;
> > >       u64 mask;
> > > +     int ret;
> > >
> > >       channel = kzalloc(sizeof(*channel), GFP_KERNEL);
> > >       if (!channel)
> > >               return -ENOMEM;
> > >
> > > -     v4l2_fh_init(&channel->fh, vdev);
> > > -     file->private_data = &channel->fh;
> > > -     v4l2_fh_add(&channel->fh);
> > > -
> > >       init_completion(&channel->completion);
> > >
> > >       channel->dev = dev;
> > > @@ -2328,6 +2325,11 @@ static int allegro_open(struct file *file)
> > >                       V4L2_CID_MIN_BUFFERS_FOR_OUTPUT,
> > >                       1, 32,
> > >                       1, 1);
> > > +     if (handler->error != 0) {
> > > +             ret = handler->error;
> > > +             goto error;
> > > +     }
> > > +
> > >       channel->fh.ctrl_handler = handler;
> > >
> > >       channel->mcu_channel_id = -1;
> > > @@ -2341,7 +2343,21 @@ static int allegro_open(struct file *file)
> > >       channel->fh.m2m_ctx = v4l2_m2m_ctx_init(dev->m2m_dev, channel,
> > >                                               allegro_queue_init);
> > >
> > > +     if (IS_ERR(channel->fh.m2m_ctx)) {
> > > +             ret = PTR_ERR(channel->fh.m2m_ctx);
> > > +             goto error;
> > > +     }
> > > +
> > > +     v4l2_fh_init(&channel->fh, vdev);  
> >
> > This call sets channel->fh.ctrl_handler to vdev->ctrl_handler, which
> > has previously been overriden by the driver to handler. Therefore, this
> > patch breaks all controls. I think we should initialize channel->fh
> > before setting any fields of this struct.
> >  
> 
> I'm not very clear about this issue.
> In my second version, Hans replied that init could be moved before return 0.
> I have sent this mail to him.

This driver uses its own v4l2_ctrl_handler, which is initialized in this
function (see the "handler" variable). After the initialization, the
handler is added to the fh as

	channel->fh.ctrl_handler = handler;

to override the ctrl_handler of the fh.

However, the v4l2_fh_init() function re-initializes the ctrl_handler of
the fh

	/* Inherit from video_device. May be overridden by the driver. */
	fh->ctrl_handler = vdev->ctrl_handler;

which reverts the ctrl_handler override by the driver.

Therefore, the ctrl_handler must be overridden after the call to
v4l2_fh_init().

Michael

> 
> > Michael
> >  
> > > +     file->private_data = &channel->fh;
> > > +     v4l2_fh_add(&channel->fh);
> > > +
> > >       return 0;
> > > +
> > > +error:
> > > +     v4l2_ctrl_handler_free(handler);
> > > +     kfree(channel);
> > > +     return ret;
> > >  }
> > >
> > >  static int allegro_release(struct file *file)  
>