Re: Re: [PATCH 5/7] media: cedrus: h265: Add a couple of error checks

[Ezequiel Garcia]

Hi Jernej,

On Tue, Jul 12, 2022 at 11:25:00PM +0200, Jernej Škrabec wrote:
> Hi Ezequiel.
> 
> Dne ponedeljek, 11. julij 2022 ob 23:31:11 CEST je Ezequiel Garcia napisal(a):
> > On Mon, Jun 20, 2022 at 07:55:15PM +0200, Jernej Skrabec wrote:
> > > Now that we have infrastructure for reporting errors, let's add two
> > > checks, which will make sure slice can be actually decoded.
> > > 
> > > Signed-off-by: Jernej Skrabec <jernej.skrabec@xxxxxxxxx>
> > > ---
> > > 
> > >  drivers/staging/media/sunxi/cedrus/cedrus_h265.c | 8 ++++++++
> > >  1 file changed, 8 insertions(+)
> > > 
> > > diff --git a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> > > b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c index
> > > cfde4ccf6011..99020b9f9ff8 100644
> > > --- a/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> > > +++ b/drivers/staging/media/sunxi/cedrus/cedrus_h265.c
> > 
> > Now that you've allowed setup to fail, I would suggest
> > to have some documentation/comments on struct cedrus_dec_ops,
> > to set the expectation/rules for each ops, including the
> > call paths for each operation, which of them are allowed to sleep,
> > etc.
> 
> Documentation can be always added, but it should be separate patch.
> 

Of course, this wasn't a comment meant to block the patch in any way.

> > 
> > > @@ -435,9 +435,17 @@ static int cedrus_h265_setup(struct cedrus_ctx *ctx,
> > > struct cedrus_run *run)> 
> > >  	 * instead of start of slice data. Padding is 8 bits at most (one 
> bit
> > >  	 set to 1 and * at most seven bits set to 0), so we have to 
> inspect
> > >  	 only one byte before slice data. */
> > > 
> > > +
> > > +	if (slice_params->data_byte_offset == 0)
> > > +		return -EOPNOTSUPP;
> > > +
> > 
> > AFAICS, cedrus_h265_setup is called from .device_run.
> > We've been discussing control validation before, and I think the
> > ideal place to do that is v4l2_ctrl_ops.s_ctrl, if that's
> > at all possible.
> 
> Yeah, this particular check can be moved to s_ctrl handler.
> 
> > 
> > Driver's mem2mem device_run are executed in the context
> > of a work_struct and the failure won't really get reported
> > up the stack.
> 
> Well, at least there will be a notice in dmesg. Not ideal, I know.
> 
> > 
> > >  	padding = (u8 *)vb2_plane_vaddr(&run->src->vb2_buf, 0) +
> > >  	
> > >  		slice_params->data_byte_offset - 1;
> > > 
> > > +	/* at least one bit must be set in that byte */
> > > +	if (*padding == 0)
> > > +		return -EINVAL;
> > > +
> > 
> > Maybe this is something to check at cedrus_buf_prepare(),
> > when the buffer is queued?
> 
> I don't think so. This check is HEVC specific, but cedrus_buf_prepare() is not 
> and we need to have slice control ready, which I'm not sure is the case for 
> cedrus_buf_prepare().
> 

Hm... this is indeed a special case. The buffer contents
depend on a control, and both buffer and control are part of the same
request.

At least the decoding job would fail, and would get signaled
by the dequeued CAPTURE buffers and the V4L2_BUF_FLAG_ERROR flag,
which makes perfect sense.

In general vb2_ops.buf_prepare() is called for each
buffer (either through VIDIOC_QBUF or through MEDIA_REQUEST_IOC_QUEUE),
so for some buffer validations .buf_prepare() _might_ work better 
than .device_run for validation.

In any case, the patch looks good, I don't think these
comments should block the patch:

Reviewed-by: Ezequiel Garcia <ezequiel@xxxxxxxxxxxxxxxxxxxx>

Thanks,
Ezequiel

> Best regards,
> Jernej
> 
> > 
> > Thanks,
> > Ezequiel
> > 
> > >  	for (count = 0; count < 8; count++)
> > >  	
> > >  		if (*padding & (1 << count))
> > >  		
> > >  			break;
> > > 
> > > --
> > > 2.36.1
> 
>