Re: [PATCH 7/7] [media] gspca: fix a v4l2-compliance failure during read()

[Hans de Goede <hdegoede@redhat.com>]

Hi,

On 09-03-16 17:03, Antonio Ospite wrote:
> v4l2-compliance fails with this message:
>
>    fail: v4l2-test-buffers.cpp(512): Expected EBUSY, got 22
>    test VIDIOC_REQBUFS/CREATE_BUFS/QUERYBUF: FAIL
>
> Looking at the v4l2-compliance code reveals that this failure is about
> the read() callback.
>
> In gspca, dev_read() is calling vidioc_dqbuf() which calls
> frame_ready_nolock() but the latter returns -EINVAL in a case when
> v4l2-compliance expects -EBUSY.
>
> Fix the failure by changing the return value in frame_ready_nolock().
>
> Signed-off-by: Antonio Ospite <ao2@ao2.it>
> ---
>   drivers/media/usb/gspca/gspca.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/media/usb/gspca/gspca.c b/drivers/media/usb/gspca/gspca.c
> index 915b6c7..de7e300 100644
> --- a/drivers/media/usb/gspca/gspca.c
> +++ b/drivers/media/usb/gspca/gspca.c
> @@ -1664,7 +1664,7 @@ static int frame_ready_nolock(struct gspca_dev *gspca_dev, struct file *file,
>   		return -ENODEV;
>   	if (gspca_dev->capt_file != file || gspca_dev->memory != memory ||
>   			!gspca_dev->streaming)
> -		return -EINVAL;
> +		return -EBUSY;
>
>   	/* check if a frame is ready */
>   	return gspca_dev->fr_o != atomic_read(&gspca_dev->fr_i);

I'm not sure that this is the right fix:


1) !gspca_dev->streaming should result in -EINVAL, this is the same as what videobuf2 is doing
2) gspca_dev->memory != memory should result in -EINVAL
3) gspca_dev->capt_file != file means calling dqbuf without having done reqbufs (through the same fd)
    which certainly seemes like -EINVAL to me.

The actual problem is that dev_read() is not catching that mmap is being in use:

static ssize_t dev_read(struct file *file, char __user *data,
                     size_t count, loff_t *ppos)
{
...
         if (gspca_dev->memory == GSPCA_MEMORY_NO) { /* first time ? */
                 ret = read_alloc(gspca_dev, file);
                 if (ret != 0)
                         return ret;
         }

It will skip the read_alloc since gspca_dev->memory is USERPTR or MMAP
and then do a dqbuf with memory == GSPCA_MEMORY_READ, triggering the
gspca_dev->memory != memory check.

There are a couple of issues here:

1) gspca_dev->memory check without holding usb_lock, the taking and
releasing of usb_lock should be moved from read_alloc() into dev_read()
itself.

2) dev_read() should not assume that reading is ok if
    gspca_dev->memory == GSPCA_MEMORY_NO, it needs a:

if (gspca_dev->memory != GSPCA_MEMORY_NO &&
     gspca_dev->memory != GSPCA_MEMORY_READ)
     return -EBUSY;

(while holding the usb_lock so the above is wrong)

3) If gspca_dev->memory == GSPCA_MEMORY_READ already the
    stream could have been stopped. so we need to check
    gspca_dev->streaming (while holding the usb_lock)
    and do a streamon if it is not set (and then we can
    remove the streamon from read_alloc())

So TL;DR: dev_read needs some love.

Regards,

Hans


p.s.

If you've time to work on v4l2 stuff what gspca really needs
is to just have its buffer handling ripped out and be rewritten
to use videobuf2. I would certainly love to see a patch for that.