Hi Hans
On Tue, 2 Oct 2012, Hans Verkuil wrote:
> On Mon 1 October 2012 19:51:18 Guennadi Liakhovetski wrote:
> > Hi Hans
> >
> > Thanks for the review. As you might have seen, I just posted v2 of this
> > driver. In it I addressed almost all your comments. As for the rest:
> >
> > On Tue, 11 Sep 2012, Hans Verkuil wrote:
> >
> > > On Tue 11 September 2012 15:01:19 Guennadi Liakhovetski wrote:
> >
> > [snip]
> >
> > > > +struct sh_veu_dev;
> > > > +
> > > > +struct sh_veu_file {
> > >
> > > struct v4l2_fh is missing here: needed to implement control events.
> >
> > I removed ctrls completely instead. Don't think mine were useful for
> > anything.
> >
> > [snip]
> >
> > > > +static int sh_veu_try_fmt(struct v4l2_format *f, const struct
> > > > sh_veu_format *fmt)
> > > > +{
> > > > + struct v4l2_pix_format *pix = &f->fmt.pix;
> > > > + unsigned int y_bytes_used;
> > > > +
> > > > + /*
> > > > + * V4L2 specification suggests, that the driver should correct
> > > > the
> > > > + * format struct if any of the dimensions is unsupported
> > > > + */
> > > > + switch (pix->field) {
> > > > + case V4L2_FIELD_ANY:
> > > > + pix->field = V4L2_FIELD_NONE;
> > >
> > > Add a 'break' here.
> >
> > That's a matter of taste, I think:-) The logic here is, that after setting
> > the field to NONE the "rest" should be identical with the case, where NONE
> > is already set by the user, so, just fall through.
>
> Can you add a /* fall through */ comment in that case? That clarifies that it
> is intentional.
>
> >
> > [snip]
> >
> > > > +static void sh_veu_colour_offset(struct sh_veu_dev *veu, struct
> > > > sh_veu_vfmt *vfmt)
> > > > +{
> > > > + /* dst_left and dst_top validity will be verified in CROP /
> > > > COMPOSE */
> > > > + unsigned int left = vfmt->frame.left & ~0x03;
> > > > + unsigned int top = vfmt->frame.top;
> > > > + dma_addr_t offset = ((left * veu->vfmt_out.fmt->depth) >> 3) +
> > > > + top * veu->vfmt_out.bytesperline;
> > > > + unsigned int y_line;
> > > > +
> > > > + vfmt->offset_y = offset;
> > > > +
> > > > + switch (vfmt->fmt->fourcc) {
> > > > + case V4L2_PIX_FMT_NV12:
> > > > + case V4L2_PIX_FMT_NV16:
> > > > + case V4L2_PIX_FMT_NV24:
> > > > + y_line = ALIGN(vfmt->frame.width, 16);
> > > > + vfmt->offset_c = offset + y_line * vfmt->frame.height;
> > > > + break;
> > > > + case V4L2_PIX_FMT_RGB332:
> > > > + case V4L2_PIX_FMT_RGB444:
> > > > + case V4L2_PIX_FMT_RGB565:
> > > > + case V4L2_PIX_FMT_BGR666:
> > > > + case V4L2_PIX_FMT_RGB24:
> > > > + vfmt->offset_c = 0;
> > > > + break;
> > > > + default:
> > > > + BUG();
> > >
> > > Wouldn't WARN_ON be more polite?
> >
> > This "default" can be entered only if someone modifies the driver and does
> > this wrongly, so, I just make sure, that the author realises their mistake
> > before shipping to the user;-)
> >
> > [snip]
> >
> > > > +static int sh_veu_s_input(struct file *file, void *fh, unsigned int i)
> > > > +{
> > > > + return i ? -EINVAL : 0;
> > > > +}
> > > > +
> > > > +static int sh_veu_g_input(struct file *file, void *fh, unsigned int *i)
> > > > +{
> > > > + *i = 0;
> > > > + return 0;
> > > > +}
> > > > +
> > > > +static int sh_veu_enum_input(struct file *file, void *fh,
> > > > + struct v4l2_input *inp)
> > > > +{
> > > > + return inp->index ? -EINVAL : 0;
> > > > +}
> > >
> > > Why implement the input ioctls at all? I'm not sure whether they are
> > > relevant here. If you do, then enum_input really has to fill in the
> > > other fields of struct v4l2_input as well. But I would just remove
> > > support for these ioctls.
> >
> > Right, I don't need them, but in the beginning I tried using gstreamer for
> > testing, and, I think, it required them... In the end I anyway gave up on
> > it and used my own test program, so, yeah, can remove them...
> >
> > [snip]
> >
> > > Please run v4l2-compliance (the latest version from v4l-utils.git) over
> > > this
> > > driver. Most if not all of the issues I've highlighted above would have
> > > been
> > > found by v4l2-compliance.
> >
> > As I also mentioned in v2, I think, the spec should be extended to also
> > allow other errors from REQBUFS.
>
> EBUSY should be documented. Perhaps you can make a patch for that? :-)
...and ENOMEM too. And then I don't see why EPERM shouldn't be allowed
either;-)
> BTW, in the second version of your patch you return -EPERM if the filehandle
> was set up for streaming.
You mean if a different filehandle has been assigned as the stream owner,
yes.
> But that isn't right. It is REQBUFS (or CREATE_BUFS
> or read/write) that sets up a filehandle for streaming. Only after calling one
> of those functions is everything locked into place. Until that time it is
> perfectly valid to change formats.
Hm, S_FMT? If one process allocated (and possibly queued) buffers and
another one is trying to set a format, for which buffers wouldn't be
suitable any more? For this reason I also consider S_FMT to be a
"privileged" operation. With CREATE_BUFS we can support several formats
without re-initialising the queue, but still, since S_FMT is so closely
related to buffer allocation, I reserve it too for the stream "owner."
This is summarised in this comment:
/*
* It is not unusual to have video nodes open()ed multiple times. While some
* V4L2 operations are non-intrusive, like querying formats and various
* parameters, others, like setting formats, starting and stopping streaming,
* queuing and dequeuing buffers, directly affect hardware configuration and /
* or execution. This function verifies availability of the requested interface
* and, if available, reserves it for the requesting user.
*/
> If an app want to get exclusive access and prevent anyone else from messing
> with formats, then it should use the priority ioctls.
Yes, I looked at those. But if they are not used, we still put some policy
in place, right? Say, we don't allow queuing and dequeuing of buffers via
2 different file-handles.
Thanks
Guennadi
---
Guennadi Liakhovetski, Ph.D.
Freelance Open-Source Software Developer
http://www.open-technology.de/
--
To unsubscribe from this list: send the line "unsubscribe linux-media" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html